Sunspot counting method and system based on deep learning

Abstract

The present application relates to a deep learning-based sunspot counting method and system, image data of the solar photosphere and corresponding time period magnetic field active region data are obtained and preprocessed, the magnetic field active region data is used for positioning processing to determine the region of interest, a binary mask image is obtained through segmentation, after error correction and reconstruction, a pseudo label corresponding to historical data is obtained, together with the preprocessed image data as the training data of the segmentation model, to obtain the reconstructed mask image; further, the no-half-shadow sunspot in the image is filtered out through the pixel intensity histogram, the part retaining the half-shadow sunspot is subjected to binary classification clustering, and the number of clusters is counted as the number of single sunspots; the reconstructed mask image is subjected to image recognition based on historical data to obtain the number of sunspot groups; the number of sunspot groups and the number of single sunspots are used to obtain the final sunspot counting result by using the Wolf number function model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sunspot counting, in particular to a sunspot counting method and system based on deep learning. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] As a visual representation of the strong magnetic field in the photosphere, the number and distribution of sunspots are the core indicators of the intensity of solar activity. The traditional relative sunspot number R = K(10g + s) quantifies the intensity of solar activity by statistically weighting the number of sunspot groups g and the number of individual sunspots s. However, this model has long relied on manual annotation by ground-based observation stations, facing technical bottlenecks: first, the image resolution is limited by atmospheric turbulence, and the characteristics of small sunspots are easily obscured by noise; second, the subjective determination criteria of different observers cause sunspot group annotation bias, making it difficult to ensure consistency of historical data; third, manual annotation of single full-disk images is time-consuming and cannot meet the real-time needs of space weather forecasting.

[0004] Currently, although high-resolution observation data outside the atmosphere can be provided by space-based telescopes, existing automated analysis methods still have significant shortcomings.

[0005] Traditional image processing methods (such as morphological operations and region growing algorithms) rely on manual threshold setting and are not adaptable to the multi-scale structure and dynamic evolution characteristics of sunspots. Although deep learning technology based on convolutional neural networks has made progress in the field of sun image segmentation, existing research has focused on single-task recognition (such as sunspot existence classification or region segmentation), and has not established an analysis framework that meets the Wolf number calculation standard, especially lacking physical constraint mechanisms in key areas such as sunspot group detection, umbra, and penumbra differentiation. SUMMARY

[0006] To solve the technical problems in the background art, the present application provides a sunspot counting method and system based on deep learning, which uses historical data combined with image segmentation algorithms to create segmentation pseudo-labels and train sunspot segmentation models, fundamentally improving annotation efficiency. The image segmentation algorithm is combined with the U-Net model to achieve robust edge segmentation and effective segmentation under weak contrast; it can automatically run to support batch processing of full-disk sequence images, reducing reliance on manpower.

[0007] To achieve the above purpose, the present application adopts the following technical solutions:

[0008] The first aspect of the present application provides a sunspot counting method based on deep learning, comprising the following steps:

[0009] Obtain image data of the solar photosphere and corresponding time period of magnetic field active region data, and pre-process;

[0010] The pre-processed image data is subjected to positioning processing using the magnetic field active region data, to determine a region of interest, to obtain a binary mask image through segmentation, and to perform error correction using historical data, to obtain a pseudo-label corresponding to the historical data after reconstruction; the obtained pseudo-label, together with the pre-processed image data, is used as training data for a segmentation model, to obtain a reconstructed mask image;

[0011] The reconstructed mask image is subjected to filtering of non-penumbral sunspots in the image through a pixel intensity histogram, and counting of the non-penumbral sunspots; the part retaining the penumbral sunspots is subjected to binary classification clustering, and the sum of the number of umbrae and the number of non-penumbral sunspots is the number of individual sunspots;

[0012] The reconstructed mask image is subjected to image recognition based on historical data, to obtain the number of sunspot groups;

[0013] The obtained number of sunspot groups and the number of individual sunspots are used to obtain the final sunspot count result using a Wolf number function model.

[0014] Further, the binary mask image obtained through segmentation is as follows: using a MorphACWE segmentation algorithm, based on the Gaussian distribution of light intensity in a quiet region of the sun, an intensity threshold for detecting sunspots is set, to obtain the binary mask image.

[0015] Further, error correction is performed using historical data, specifically: sunspot group annotation data in the historical data is extracted and converted into a binary mask, and through mask point multiplication operation, sunspots that do not meet the set requirements in the magnetic field active region are excluded, to perform error correction on the segmentation result of MorphACWE.

[0016] Further, reconstruction is performed to obtain a pseudo-label corresponding to the historical data, specifically: according to the cosine law of the sun's spherical coordinates and a set minimum sunspot area threshold, the apparent pixel size of the smallest sunspot at a certain position in the mask image is determined If there is a region in the sunspot detected by the MorphACWE algorithm with an area less than , it is considered as a false detection and is excluded, to obtain the pseudo-label corresponding to the historical data.

[0017] Further, the reconstructed mask image is subjected to filtering of non-penumbral sunspots in the image through a double-peak structure in the pixel intensity histogram, to obtain the number of non-penumbral sunspots and the part retaining the penumbral sunspots;

[0018] The part of the sunspot remaining the penumbra is subjected to the binary classification clustering of "the umbra vs. the background + the penumbra", the number of the umbra clustered is counted as the sunspot number containing the penumbra, and is added to the sunspot number without the penumbra to obtain the single sunspot number.

[0019] Further, the mask image after reconstruction is subjected to image recognition based on historical data to obtain the sunspot group number, specifically: the background of the mask image after reconstruction is filled with the average intensity of the non-sunspot region, and the sunspot group boundary box labeled according to the historical data is taken as a label and input into a YOLO model to obtain the sunspot group number.

[0020] Further, the Wolf number function model is used to obtain the final sunspot counting result, specifically: the total sunspot number is calculated according to R = S + 10·G, wherein S is the single sunspot number obtained, and G is the sunspot group number obtained.

[0021] The second aspect of the present application provides a sunspot counting system based on deep learning, comprising:

[0022] The data acquisition and preprocessing module is configured to: acquire image data of the photosphere of the sun and magnetic field active region data of a corresponding period, and pre-process the data;

[0023] The pseudo-label generation module is configured to: the pre-processed image data is subjected to positioning processing using the magnetic field active region data to determine a region of interest, a binary mask image is obtained through segmentation, and error correction is performed using historical data, and after reconstruction, a pseudo-label corresponding to the historical data is obtained;

[0024] The U-Net model training module is configured to: the obtained pseudo-label is taken together with the pre-processed image data as training data of a segmentation model to obtain a mask image after reconstruction;

[0025] The single sunspot number module is configured to: the mask image after reconstruction is subjected to filtering of the sunspot without the penumbra through a pixel intensity histogram, and the sunspot without the penumbra is counted; the part of the sunspot remaining the penumbra is subjected to binary classification clustering, and the sum of the number of the umbra and the number of the sunspot without the penumbra is the single sunspot number;

[0026] The sunspot group detection module is configured to: the mask image after reconstruction is subjected to image recognition based on historical data to obtain the sunspot group number;

[0027] The total number calculation module is configured to: the sunspot group number and the single sunspot number are subjected to the Wolf number function model to obtain the final sunspot counting result.

[0028] The third aspect of the present application provides a computer program product comprising computer readable instructions which, when run on an electronic device, cause the electronic device to implement the deep learning-based sunspot counting method described above.

[0029] The fourth aspect of the present application provides an electronic device comprising at least one processor and a memory connected to the processor, the memory being configured to store a computer program; the processor being configured to execute the computer program, so that the electronic device can implement the deep learning-based sunspot counting method described above.

[0030] Compared with the prior art, the above one or more technical solutions have the following beneficial effects:

[0031] 1. By comprehensively applying image segmentation models for sunspot segmentation, clustering, and distinguishing umbra and penumbra, and image recognition models for sunspot group detection, the image processing process is integrated with solar physics observation standards, and the systematic upgrade from scientific logic to technical path is realized.

[0032] 2. Since sunspots always appear in active regions, the full-disk white light image corresponding to the magnetic field active region is cut out to avoid noise in the quiet sun region being identified as sunspots. Furthermore, by using image segmentation methods to create segmentation pseudo-labels and train sunspot segmentation models, manual annotation can be replaced, and annotation efficiency can be fundamentally improved.

[0033] 3. During the generation of pseudo-labels, the adaptive threshold based on light intensity Gaussian distribution is better than the fixed threshold, which can detect weak contrast sunspots; in combination with historical data to correct pseudo-labels, non-physical noise is filtered to improve reliability; in combination with the cosine law of the sun's coordinates to dynamically adjust the minimum sunspot area threshold, the result is ensured to be reasonable in the physical aspect.

[0034] 4. The morphological segmentation model (MorphACWE) is introduced to cooperate with the deep learning image segmentation method to achieve robust edge segmentation and effective segmentation under weak contrast.

[0035] 5. During counting, considering that umbra can exist independently while penumbra cannot, in combination with historical data annotation of group boundary boxes for training, the sunspot group position and number are directly outputted to avoid manual definition of group merging rules (such as distance threshold).

[0036] 6. In view of the problem that "no analysis framework meeting the Wolf number calculation standard has been established", the present solution not only constructs an analysis framework, but also aligns it with historical observation data to enhance the explainability and standardization degree of the automatic identification result.

[0037] 7. To solve the problem of "lack of physical constraint mechanism", the scheme introduces magnetic field active region data as the physical boundary reference of the active area, which is used for the extraction of the region of interest; edge darkening correction is performed on the input image to improve the segmentation effect of the near-solar disk edge region; the longitude and latitude of the sunspot are considered when false sunspots are generated, and the number of pixels occupied by the false sunspots is calculated; the YOLO model is used for group detection, and the sunspot group division in the historical data is used as the training label, so that the division standard is closer to the human division standard based on the spatial structure and sunspot distribution. BRIEF DESCRIPTION OF DRAWINGS

[0038] The drawings accompanying the specification of this application form a part thereof, serve to further provide a further understanding of the application, and together with the description of the exemplary embodiments of the application, serve to explain the application, and do not constitute an improper limitation of the application.

[0039] Figure 1 is a schematic diagram of the overall process of the deep learning-based sunspot counting method provided by one or more embodiments of the application;

[0040] Figure 2 is a schematic diagram of the pseudo-label generation process during sunspot counting provided by one or more embodiments of the application;

[0041] Figure 3 is a schematic diagram of the sunspot segmentation process during sunspot counting provided by one or more embodiments of the application;

[0042] Figure 4 is a schematic diagram of the sunspot counting provided by one or more embodiments of the application;

[0043] Figure 5 is a schematic diagram of the correlation comparison results of the estimated total sunspot number and SIDC-SILSO, USET provided by one or more embodiments of the application;

[0044] Figure 6 is a schematic diagram of the daily comparison results of the estimated total sunspot number and SIDC-SILSO provided by one or more embodiments of the application;

[0045] Figure 7 is a schematic diagram of the daily comparison results of the estimated total sunspot number and USET provided by one or more embodiments of the application;

[0046] Figure 8 is a schematic diagram of the correlation comparison results of the estimated total sunspot number and SWPC, DPD provided by one or more embodiments of the application;

[0047] Figure 9 is a schematic diagram of the daily comparison of the estimated total sunspot number and SWPC provided by one or more embodiments of the application;

[0048] Figure 10 is a daily comparison between the total sunspot number estimated and the DPD provided by one or more embodiments of the present application;

[0049] Figure 11 is a schematic diagram of a verification experiment result of solar activity periodicity (data during the period from July 2010 to June 2023) provided by one or more embodiments of the present application;

[0050] Figure 12 is a schematic diagram of a verification experiment result of solar activity periodicity (test set data) provided by one or more embodiments of the present application. DETAILED DESCRIPTION

[0051] The present application will be further described below in conjunction with the accompanying drawings and embodiments.

[0052] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0053] Sunspots usually occur in groups, and an active region may contain multiple sunspots, while isolated sunspots are less active. Sunspot groups are usually accompanied by stronger magnetic field activity and are more likely to trigger flares and coronal mass ejections (CMEs), so the Wolf Number is used for sunspot observation to reflect the "quality" and "quantity" of solar activity through weighted calculation, which is an international standard index for sunspot activity.

[0054] As introduced in the background art, conventional image processing methods (such as morphological operations and region growing algorithms) rely on manual threshold setting and are not adaptable to the multi-scale structure and dynamic evolution characteristics of sunspots. Although deep learning technology based on convolutional neural networks has made progress in the field of solar image segmentation, existing researches mostly focus on single-task recognition (such as sunspot existence classification or region segmentation), and have not established an analysis framework consistent with the Wolf Number calculation standard, especially lacking physical constraint mechanisms in key links such as sunspot group detection, umbra, and penumbra differentiation.

[0055] Regarding the "lack of an analysis framework consistent with the Wolf number calculation standard", the current sunspot count is mainly based on the Wolf number (R = k(10g + s)), where g is the number of sunspot groups, s is the total number of sunspots, and k is the correction coefficient. Most existing research focuses on a single task (such as sunspot segmentation or detection), which can detect the number of sunspots in the image, but lacks a unified framework to ensure that the extracted "sunspot groups" and "individual sunspots" are consistent with the Wolf number standard defined in historical manual observations, resulting in a significant deviation between the automatic estimate and the official statistical results (such as SIDC or NOAA).

[0056] Regarding the "lack of physical constraint mechanism", the following points are mainly referred to:

[0057] The physical boundary of the solar active region is not clear, and most algorithms are based on image feature segmentation, ignoring the influence of magnetic field structure on sunspot distribution;

[0058] Sunspot segmentation lacks processing of optical effects such as limb darkening, affecting the recognition of edge sunspots;

[0059] Sunspot group classification ignores spatial continuity, life cycle, and other physical characteristics.

[0060] The following embodiments give a deep learning-based sunspot counting method and system, which innovatively combines the MorphACWE method with the U-Net segmentation network, and combines the segmentation results with the sunspot annotation data provided by the database (USET) provided by the Royal Observatory of Belgium, to ensure the consistency of the segmentation results with the ground historical observation standard. By constructing a sun full-disk image sunspot group detection system based on the YOLO model, combined with a sunspot umbra counting algorithm based on pixel intensity and K-Means, accurate counting of individual sunspots and sunspot groups is achieved. This method effectively overcomes the limitations of ground observation data quality, low efficiency of manual annotation, and other defects, providing core technical support for the construction of an automated calibration system for a cross-century sunspot database.

[0061] Embodiment One:

[0062] This scheme obtains full-disk continuous spectral observation data of the Solar Dynamics Observatory (SDO) Helioseismic and Magnetic Imager (HMI) from the Joint Science Operations Center (JSOC) database.

[0063] Since SDO / HMI has been regularly observed since 2010, this scheme uses data from July 2010 to June 2023 as the data source for this scheme.

[0064] The data preprocessing procedure includes the de-rotation correction of the solar north-south axial inclination and the correction of the edge darkening effect. The image resolution used is 4096x4096 pixels to ensure the detection capability of the smallest sunspot (usually represented by only a few pixels).

[0065] To avoid potential data leakage problems, part of the data is deliberately excluded in this scheme. The training set is constructed using observation data from January 1st to September 15th of each year, and the test set is composed of data from October 1st to December 15th of each year.

[0066] To further select the sunspot-containing solar region as the training set, part of the data without sunspots is excluded, and the final training set contains 2464 images, and the test set covers 775 images.

[0067] At the same time, this scheme obtains HMI active region patch data (HARP) from the JSOC database for the positioning of active regions in white light images. This dataset provides approximate boundary information of complete sunspot groups, which becomes an effective tool for selecting interesting regions in full-disk observation data. Both the full-disk continuum spectral observation data and the HARP active region data are downloaded synchronously at 12:00 every day in international atomic time to ensure the time consistency of the two data sets.

[0068] In addition, based on the manual annotation results of sunspots from the USET dataset of the Royal Observatory of Belgium solar equator table, the sunspot groups are manually annotated: for each SDO / HMI white light image, the corresponding USET annotation is matched, and high-precision annotation of sunspot groups is achieved by minimizing the bounding boxes. These annotation results are used to optimize the segmentation results of the segmentation model, and also serve as the training benchmark for the target detection model, thereby improving the detection accuracy of the automated framework.

[0069] As shown in Figure 1 , the deep learning-based sunspot counting method includes the following steps:

[0070] Obtain image data of the solar photosphere and corresponding period magnetic field active region data, and preprocess;

[0071] The preprocessed image data is positioned using the magnetic field active region data to determine the region of interest, and a binary mask image is obtained by segmentation. Error correction is performed using historical data, and after reconstruction, a pseudo-label corresponding to the historical data is obtained. The obtained pseudo-label together with the preprocessed image data is used as training data for the segmentation model, and a reconstructed mask image is obtained;

[0072] The reconstructed mask image filters out the sunspot-free penumbra in the image through a pixel intensity histogram and counts the sunspot-free penumbra; the part retaining the penumbra is subjected to binary classification clustering, and the sum of the umbra number and the sunspot-free penumbra number obtained by clustering is the number of single sunspots;

[0073] The reconstructed mask image is subjected to image recognition based on historical data to obtain the number of sunspot groups;

[0074] The obtained number of sunspot groups and the number of single sunspots are subjected to Wolf number function modeling to obtain the final sunspot counting result.

[0075] Step 1: Pseudo-label generation and model training.

[0076] S1.1 Pseudo-label generation, as shown in Figure 2 The present embodiment takes the HARP region of NOAA 13186 as an example.

[0077] The original image is obtained and preprocessed, the SDO / HMI white light image is subjected to edge darkening correction and inclination rotation correction, the HARP data is used to locate the active region coordinates, the corrected image is determined, and the region of interest is obtained.

[0078] MorphACWE segmentation: based on the Gaussian distribution (mean μ and standard deviation σ) of the light intensity in the quiet region of the sun, the intensity threshold for detecting sunspots is set to 5 times the standard deviation (5σ), the iteration number of the MorphACWE algorithm is 100 times, and the minimum pixel number threshold is 1, to generate a preliminary binary mask.

[0079] Historical data calibration: the sunspot group label made according to the USET data set of the Royal Observatory of Belgium is converted into a binary mask, and the non-USET-defined small sunspots in the HARP active region are excluded through mask point multiplication operation, and the MorphACWE segmentation result is subjected to error correction.

[0080] Spatial resolution filtering: considering that the spatial resolution of the SDO / HMI imaging system is 0.505 arcsec / pixel, which corresponds to about 362.6 km / pixel at the center of the sun, and the smallest observable sunspot diameter is about 1500 km, which covers about 4.14 pixels after conversion. To ensure the physical reasonableness of the segmentation result, 4.14 pixels are set as the minimum sunspot area threshold in the central region of the sun. In addition, however, due to the projection deformation effect, the apparent size of the sunspot near the edge of the sun will decrease, and this deformation follows the cosine law of the sun coordinate.

[0081] Specifically, the apparent pixel size of the smallest sunspot at a certain position can be represented as:

[0082] ​ ;

[0083] wherein, and denote the sun's latitude and longitude at that location, respectively.

[0084] Therefore, if there is a region in the sunspot detected by the MorphACWE algorithm with an area smaller than the minimum sunspot size, it is considered a false detection and is removed. After the above spatial resolution filtering (i.e., full-disk reconstruction), the final output is a pseudo-label that aligns with the SIDC-SILSO historical record.

[0085] The MorphACWE algorithm is an image segmentation algorithm based on morphological operations, which is an improved version of the traditional ACWE (Active Contours Without Edges) model. It contains a morphological dilation / erosion step to replace the curvature term in the traditional ACWE, as well as a region mean calculation similar to the region statistics term in ACWE, and a threshold processing step to update the contour based on the region statistics information. The MorphACWE algorithm achieves image segmentation by alternating the use of these steps.

[0086] In this embodiment, the MorphACWE algorithm is used to generate pseudo-labels, train the U-Net segmentation model, and the final segmentation result is provided by the U-Net. Since MorphACWE is good at handling edge-blurred segmentation cases, it only requires the average pixel value of the target to be segmented and the background to be different, and does not require the target to have a clear edge, making it suitable for segmentation processing based on the sunspot morphology in the image.

[0087] S1.2 U-Net model training, as shown in Figure 3 .

[0088] Data loading: Use the DataLoader class of the PyTorch framework for data loading and batch processing. Through a custom Dataset class, image data and pseudo-labels are packaged into a unified format, and a data loader is defined for batch processing. The functions of the data loader include automatically loading images and labels, data normalization and standardization, and inputting the model by batch.

[0089] Chunk processing strategy: Divide the full-disk image (4096x4096) into 16 1024x1024 sub-images to input the U-Net, solving the resource occupation problem and small target detection problem.

[0090] Loss function and optimization: Use binary cross-entropy loss (BCELoss), Adam optimizer (learning rate 0.001), group normalization (GroupNorm) instead of batch normalization, and train for 1 cycle.

[0091] Output and reconstruction: After the sub-image segmentation result is filtered by spatial resolution (i.e., full-disk reconstruction), it is spliced into a full-disk probability map, and a final segmentation mask is generated by thresholding at 0.5. The mask is multiplied element by element with the preprocessed image to obtain the reconstructed image.

[0092] Step 2: Single sunspot counting, as shown in Figure 4 .

[0093] S2.1 Umbra-penumbra differentiation.

[0094] After preprocessing, the image determines the region of interest based on the mask result, detects the bimodal structure through the pixel intensity histogram, obtains the penumbra feature, filters out the sunspots without penumbra in the image, and obtains the number of sunspots without penumbra and the part containing penumbra. The part containing penumbra is clustered into two categories (umbra vs. background + penumbra), and the number of umbra clusters is counted as the number of sunspots containing penumbra, which is added to the number of sunspots without penumbra to obtain the number of single sunspots.

[0095] Among them, connected component analysis: based on mask recognition connected component, through pixel intensity histogram detection bimodal structure (penumbra feature), filter out sunspots without penumbra.

[0096] Among them, K-means clustering: two-class clustering (umbra vs. penumbra) is performed on the retained sunspot part, and the number of umbra clusters is counted as the number of single sunspots.

[0097] Step 3: Sunspot group detection and total number calculation, as shown in Figure 4 .

[0098] S3.1 YOLO sunspot group detection.

[0099] Input processing: Fill the mask image background with the average intensity of non-sunspot regions, and input the YOLOv5 model with the sunspot group bounding box labeled according to the USET label.

[0100] Model training: Scale the image to 1024x1024, batch size 9, learning rate 0.01, train for 30 cycles, and retain the original loss function.

[0101] Result output: Output the sunspot group bounding box and count the number of detected sunspot groups.

[0102] S3.2 Total number calculation.

[0103] Formula application: Calculate the total number of sunspots according to R = S + 10·G, where S (corresponding to Ns in Figure 4 ) is the number of single sunspots obtained in step S2.1, and G (corresponding to Ns in Figure 4 ).Ng is the number of sunspot groups obtained in step S3.1.

[0104] The above scheme introduces a morphological model (MorphACWE) to cooperate with the U-Net, realizes strong edge robustness, and still effective segmentation effect under weak contrast.

[0105] The above scheme can replace manual annotation: using HARP+MorphACWE to make segmentation pseudo-labels, training sunspot segmentation model, and fundamentally improving annotation efficiency.

[0106] The above scheme realizes full-process automation: supporting batch processing of full-disk sequence images, automatically outputting sunspot number, sunspot group and its statistical results, and reducing human dependence.

[0107] In view of the problem that "no analysis framework is established to meet the Wolf number calculation standard", the scheme not only constructs the analysis framework, but also aligns it with the historical observation data, enhances the explainability and standardization of the automatic identification results.

[0108] In view of the problem of "lack of physical constraint mechanism", the scheme introduces HARP data as the reference of the physical boundary of the active area for the extraction of the region of interest; performs edge darkening correction on the input image to improve the segmentation effect of the near-disk edge region; uses YOLO model for group detection, and uses USET group division as training label to make it closer to the division standard based on spatial structure and sunspot distribution by human.

[0109] By comprehensively applying the U-Net model for sunspot segmentation, K-means clustering to distinguish the umbra and penumbra, and the YOLO model to detect sunspot groups, the image processing process is integrated with the standard of solar physics observation, realizing the systematic upgrading from scientific logic to technical path.

[0110] In addition, a correction mask is constructed based on the Ucled Solar Equatorial Table (USET) dataset to make the estimated results consistent with the historical sunspot records of the past two centuries.

[0111] According to the relevant experimental results, the total number of sunspots estimated by this method has a high correlation of 0.97 with the sunspot index provided by the Solar Influences Data Analysis Center (SIDC-SILSO), verifying the robustness and reliability of the method. The proposed method provides an efficient and objective solution for sunspot number estimation, overcoming the limitations of existing manual and basic image processing methods, and has good scalability, suitable for future high-resolution solar observation data.

[0112] Experimental verification. Verify the consistency of the present scheme with historical data, the performance across datasets, and the solar activity periodicity using published existing data.

[0113] Published existing data include:

[0114] SIDC: Sunspot Index and Long-term Solar Observations of the Solar Influences Data Analysis Center, SIDC-SILSO (https: / / www.sidc.be / SILSO / DATA / SN_d_tot_V2.0.txt);

[0115] USET: Uccle Solar Equatorial Table, USET (https: / / vo-tap.oma.be / );

[0116] SWPC: Space Weather Prediction Center, SWPC (https: / / www.swpc.noaa.gov / products / solar-region-summary);

[0117] DPD: Debrecen Photoheliographic Data, DPD (http: / / fenyi.solarobs.epss.hun-ren.hu / DPD / ).

[0118] Experiment 1: Historical data consistency verification;

[0119] Comparison object: SIDC-SILSO, USET;

[0120] Dataset: Use SIDC-SILSO as the benchmark and USET as the auxiliary calibration source.

[0121] Result: The estimated daily total sunspot number is consistent with the trend of SIDC-SILSO, and the Pearson correlation coefficient reaches 0.97 (as shown in Figure 5 and Figure 6 ).

[0122] The estimated daily total sunspot number is consistent with the trend of USET, and the Pearson correlation coefficient reaches 0.93 (as shown in Figure 5 andFigure 7 As shown in FIG. 1.

[0123] Experiment 2: Performance comparison across datasets;

[0124] Comparison object: SWPC, DPD database;

[0125] Results: The estimated daily total sunspot number is consistent with the trend of SWPC, and the Pearson correlation coefficient reaches 0.94 (as shown in FIG. 2 and FIG. 3). Figure 8 and Figure 9 As shown in FIG. 2 and FIG. 3).

[0126] The estimated daily total sunspot number is consistent with the trend of DPD, and the Pearson correlation coefficient reaches 0.81 (without USET mask correction) (as shown in FIG. 4 and FIG. 5). Figure 8 and Figure 10 As shown in FIG. 4 and FIG. 5).

[0127] Experiment 3: Verification of solar activity periodicity (butterfly diagram analysis) ;

[0128] Method: Statistics of the latitude distribution of sunspot groups from 2010 to 2023, and draw the latitude-time evolution diagram.

[0129] Results: The sunspot group presents a classic “butterfly diagram” pattern (as shown in FIG. 6 and FIG. 7), that is, the sunspot group is distributed in the ±30°-40° high latitude area at the beginning of the solar cycle, and then gradually shifts to the equator. Figure 11 and Figure 12 As shown in FIG. 6 and FIG. 7).

[0130] Example Two

[0131] The sunspot counting system based on deep learning comprises:

[0132] The data acquisition and preprocessing module is configured to acquire image data of the solar photosphere and magnetic field active region data of the corresponding period, and to preprocess the data.

[0133] The pseudo-label generation module is configured to perform positioning processing on the preprocessed image data using the magnetic field active region data, determine the region of interest, obtain a binary mask image through segmentation, and perform error correction using historical data. After reconstruction, the pseudo-label corresponding to the historical data is obtained.

[0134] The U-Net model training module is configured to obtain the pseudo-label together with the preprocessed image data as training data for the segmentation model, and obtain the reconstructed mask image.

[0135] The single sunspot number module is configured to: the reconstructed mask image is filtered by a pixel intensity histogram to remove the sunspot-free penumbra in the image, and the sunspot-free penumbra is counted; the part remaining the penumbra is subjected to binary classification clustering, and the sum of the umbra number and the sunspot-free penumbra number is the single sunspot number;

[0136] The sunspot group detection module is configured to: the reconstructed mask image is subjected to image recognition based on historical data to obtain the number of sunspot groups;

[0137] The total number calculation module is configured to: the number of sunspot groups and the single sunspot number are used to obtain the final sunspot counting result by using a Wolf number function model.

[0138] The MorphACWE method is combined with the U-Net segmentation network, and the segmentation result is combined with the sunspot annotation data provided in the database (USET) provided by the Royal Observatory of Belgium, to ensure the consistency of the segmentation result with the ground historical observation standard. By constructing a sun full-disk image sunspot group detection system based on the YOLO model, and combining a sunspot umbra counting algorithm based on pixel intensity and K-Means, the accurate counting of single sunspots and sunspot groups is realized. This method effectively overcomes the defects of limited ground observation data quality and low artificial annotation efficiency, and provides core technical support for constructing an automatic calibration system for a cross-century sunspot database.

[0139] Embodiment three

[0140] A computer program product includes computer readable instructions that, when executed on an electronic device, cause the electronic device to implement the above deep learning-based sunspot counting method.

[0141] Embodiment four

[0142] An electronic device includes at least one processor and a memory connected to the processor, the memory being configured to store a computer program; the processor is configured to execute the computer program, so that the electronic device can implement the above deep learning-based sunspot counting method.

[0143] Embodiment five

[0144] A computer storage medium stores one or more computer programs, and when the one or more computer programs are executed by an electronic device, the electronic device can implement the above deep learning-based sunspot counting method.

[0145] The above merely describes the preferred embodiments of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A sunspot counting method based on deep learning, characterized by, The method comprises the following steps: obtaining image data of a solar photosphere and magnetic field active region data of a corresponding period, and preprocessing; the preprocessed image data is subjected to positioning processing by using the magnetic field active region data, a region of interest is determined, a binary mask image is obtained through segmentation, error correction is performed by using historical data, and after reconstruction, a pseudo label corresponding to the historical data is obtained; the obtained pseudo label is used as training data of a segmentation model together with the preprocessed image data, and a reconstructed mask image is obtained; the reconstructed mask image is filtered through a pixel intensity histogram to remove sunspots without penumbra, and the sunspots without penumbra are counted; the part retaining the sunspots with penumbra is subjected to binary classification clustering, and the sum of the number of umbra and the number of sunspots without penumbra is the number of single sunspots; the reconstructed mask image is subjected to image recognition based on historical data, and the number of sunspot groups is obtained; the obtained number of sunspot groups and the number of single sunspots are used to obtain the final sunspot count result by using a Wolf number function model.

2. The deep learning based sunspot counting method of claim 1, wherein, The binary mask image obtained through segmentation is specifically obtained by using a MorphACWE segmentation algorithm, setting an intensity threshold for detecting sunspots based on the Gaussian distribution of light intensity in a quiet region of the sun, and obtaining the binary mask image. 3.The deep learning based sunspot counting method of claim 1, wherein, Error correction is performed by using historical data, specifically by extracting sunspot group annotation data in the historical data, converting the sunspot group annotation data into a binary mask, excluding sunspots in the magnetic field active region that do not meet the set requirements through mask point multiplication operation, and performing error correction on the segmentation result of the MorphACWE. 4.The deep learning based sunspot counting method of claim 1, wherein, Reconstruction is obtained corresponding to the pseudo label of historical data, specifically: according to the cosine law of the day and the set minimum sunspot area threshold, the apparent pixel size of the minimum sunspot in the mask image at a certain position is determined , in the sunspot detected by MorphACWE algorithm, if the area is less than , it is considered as false detection and is removed to obtain the pseudo label corresponding to the historical data. 5.The deep learning based sunspot counting method of claim 1, wherein, The reconstructed mask image is filtered through a double-peak structure in the pixel intensity histogram to determine the penumbra feature, remove sunspots without penumbra in the image, and obtain the number of sunspots without penumbra and the part retaining sunspots with penumbra; the part retaining the sunspots with penumbra is subjected to binary classification clustering of "umbra vs. background + penumbra", the number of umbra obtained through clustering is counted as the number of sunspots with penumbra, and the number of single sunspots is obtained by adding the number of sunspots without penumbra. 6.The deep learning based sunspot counting method of claim 1, wherein, The reconstructed mask image is subjected to image recognition based on historical data, and the number of sunspot groups is obtained, specifically by filling the background of the reconstructed mask image with the average intensity of the non-sunspot region, inputting the sunspot group bounding box labeled according to the historical data as a label into a YOLO model, and obtaining the number of sunspot groups.

7. The deep learning based sunspot counting method of claim 1, wherein, The final sunspot count result is obtained by using a Wolf number function model, specifically by calculating the total number of sunspots according to R=S+10·G, wherein S is the number of single sunspots, and G is the number of sunspot groups.

8. A sunspot counting system based on deep learning, characterized in that, It comprises: a data acquisition and preprocessing module configured to obtain image data of a solar photosphere and magnetic field active region data of a corresponding period, and to preprocess the image data; a pseudo label generation module configured to perform positioning processing on the preprocessed image data by using the magnetic field active region data, to determine a region of interest, to obtain a binary mask image through segmentation, and to perform error correction by using historical data, and to obtain a pseudo label corresponding to the historical data after reconstruction; a U-Net model training module configured to use the obtained pseudo label as training data of a segmentation model together with the preprocessed image data, and to obtain a reconstructed mask image; The single sunspot number module is configured to: filter out the sunspotless sunspots in the image through a pixel intensity histogram of the reconstructed mask image, and count the sunspotless sunspots; and perform binary classification clustering on the part of the image retaining the penumbra, and the sum of the number of the umbra and the number of the sunspotless sunspots is the single sunspot number; The sunspot group detection module is configured to: perform image recognition on the reconstructed mask image based on historical data, and obtain the number of sunspot groups of the sun; The total number calculation module is configured to: utilize a Wolf number function model to obtain a final sunspot counting result from the number of the sunspot groups and the single sunspot number.

9. A computer program product, characterised in that, The computer readable instructions, when executed on the electronic device, cause the electronic device to implement the steps in the sunspot counting method based on deep learning according to any one of claims 1-7.

10. An electronic device, comprising: The computer readable instructions, when executed on the electronic device, cause the electronic device to implement the steps in the sunspot counting method based on deep learning according to any one of claims 1-7.

Images