Sunspot counting method and system based on deep learning

Abstract

The invention relates to a sunspot counting method and system based on deep learning, and the method comprises the steps: obtaining the image data of a solar photosphere layer and the data of a magnetic field activity region in a corresponding time period, carrying out the preprocessing, carrying out the positioning processing through the data of the magnetic field activity region, determining a region of interest, obtaining a binary mask image through segmentation, and carrying out the error correction and reconstruction. A pseudo label corresponding to historical data is obtained, the pseudo label and the preprocessed image data serve as training data of a segmentation model, and a reconstructed mask image is obtained; further filtering out penumbra-free black pieces in the image through a pixel intensity histogram, carrying out dichotomy clustering on a part in which penumbra black pieces are reserved, and counting the number of clustered umbra as the number of single black pieces; the reconstructed mask image is subjected to image recognition based on historical data, and the number of sunspot groups is obtained; and utilizing a Wolf number function model to obtain a final sunspot counting result according to the obtained sunspot group number and the single sunspot number.

Description

Technical Field

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

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

[0003] Sunspots, as a visual representation of the strong magnetic field in the photosphere, are a core indicator of solar activity intensity, and their number and distribution are key indicators. The traditional sunspot relative number, R=K(10g+s), constructs a classic model for quantifying solar activity intensity by weightedly counting the number of sunspot groups, g, and the number of individual sunspots, s. However, this model has long relied on manual annotation from ground-based observatories and faces technical bottlenecks: First, atmospheric turbulence limits image resolution, making tiny sunspot features easily obscured by noise; second, differences in subjective judgment criteria among observers lead to deviations in sunspot group annotation, making it difficult to ensure consistency in historical data; and third, manual annotation of a single full-disk image is time-consuming and unable to meet the real-time requirements of space weather forecasting.

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

[0005] Traditional image processing methods (such as morphological operations and region growing algorithms) rely on manual thresholding and are insufficiently adaptable to the multi-scale structure and dynamic evolution of sunspots. While deep learning techniques based on convolutional neural networks have made progress in solar image segmentation, existing research has largely focused on single-task identification (such as sunspot presence classification or region segmentation). An analytical framework that meets Wolf number calculation criteria has yet to be established. In particular, key aspects such as sunspot group detection and umbra / penumbra differentiation lack physical constraints. Summary of the Invention

[0006] To address the technical issues presented in the aforementioned background technology, the present invention provides a deep learning-based sunspot counting method and system. This method utilizes historical data combined with an image segmentation algorithm to generate segmentation pseudo-labels and train a sunspot segmentation model, fundamentally improving labeling efficiency. The image segmentation algorithm, combined with a U-Net model, achieves robust edge segmentation and effective performance even in low-contrast conditions. The system also supports automated batch processing of full-disk image sequences, reducing reliance on human labor.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions: A first aspect of the present invention provides a sunspot counting method based on deep learning, comprising the following steps: Obtain image data of the solar photosphere and magnetic field active area data of the corresponding period, and pre-process them; The preprocessed image data is positioned using magnetic field activity zone data to determine the region of interest. A binary mask image is obtained through segmentation, and error correction is performed using historical data. After reconstruction, pseudo labels corresponding to the historical data are obtained. The pseudo labels obtained, together with the preprocessed image data, serve as training data for the segmentation model to obtain a reconstructed mask image. The reconstructed mask image is filtered out of the non-penumbra sunspots in the image through the pixel intensity histogram, and the non-penumbra sunspots are counted; the part with the penumbra sunspots is clustered by binary classification, and the sum of the clustered umbra number and the non-penumbra sunspot number is the number of single sunspots; The reconstructed mask image is used for image recognition based on historical data to obtain the number of sunspot groups; The obtained sunspot group number and single sunspot number are used to obtain the final sunspot counting result using the Wolf number function model.

[0008] Furthermore, a binary mask image is obtained through segmentation. Specifically, the MorphACWE segmentation algorithm is used to set an intensity threshold for detecting sunspots based on the Gaussian distribution of light intensity in the quiet area of ​​the sun to obtain a binary mask image.

[0009] Furthermore, historical data are used for error correction. Specifically, the sunspot group annotation data in the historical data are extracted and converted into a binary mask. The sunspots that do not meet the set requirements in the magnetic field active area are eliminated through the mask dot multiplication operation, and the error of the MorphACWE segmentation result is corrected.

[0010] Furthermore, after reconstruction, the pseudo labels corresponding to the historical data are obtained. Specifically, according to the cosine law of the heliospheric coordinates and the set minimum sunspot area threshold, the apparent pixel size of the minimum sunspot at a certain position in the mask image is determined. Among the sunspots detected by the MorphACWE algorithm, if there is an area smaller than If the area is not detected, it is considered as a false detection and is removed to obtain a pseudo label corresponding to the historical data.

[0011] Furthermore, the reconstructed mask image determines the penumbra features through the bimodal structure in the pixel intensity histogram, and the non-penumbra sunspots in the image are filtered out to obtain the number of non-penumbra sunspots and the fraction of retained penumbra sunspots; The part with penumbra sunspots is clustered into two categories: "umbra vs. background + penumbra". The number of umbras obtained by clustering is counted as the number of sunspots with penumbra, and added to the number of sunspots without penumbra to obtain the number of single sunspots.

[0012] Furthermore, the reconstructed mask image is used for image recognition based on historical data to obtain the number of sunspot groups. Specifically, the background of the reconstructed mask image is filled with the average intensity of the non-spot area, and the sunspot group bounding box marked according to the historical data is used as a label and input into the YOLO model to obtain the number of sunspot groups.

[0013] Furthermore, 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, where S is the number of individual sunspots obtained and G is the number of sunspot groups obtained.

[0014] A second aspect of the present invention provides a sunspot counting system based on deep learning, comprising: The data acquisition and preprocessing module is configured to: acquire image data of the solar photosphere and data of the magnetic field active area in the corresponding period, and preprocess them; The pseudo label generation module is configured to: perform positioning processing on the pre-processed image data using the magnetic field activity area data to determine the region of interest, obtain a binary mask image through segmentation, perform error correction using historical data, and obtain a pseudo label corresponding to the historical data after reconstruction; The U-Net model training module is configured to use the pseudo labels and the preprocessed image data as training data for the segmentation model to obtain a reconstructed mask image. The single sunspot counting module is configured to: filter out the non-penumbra sunspots in the reconstructed mask image through the pixel intensity histogram and count the non-penumbra sunspots; perform binary clustering on the portion of the sunspots that retain the penumbra, and the sum of the clustered umbra number and the non-penumbra sunspot number 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 to obtain the number of sunspot groups; The total number calculation module is configured to: use the obtained sunspot group number and single sunspot number using the Wolf number function model to obtain the final sunspot counting result.

[0015] A third aspect of the present invention provides a computer program product comprising computer-readable instructions. When the computer-readable instructions are executed on an electronic device, the electronic device implements the above-mentioned sunspot counting method based on deep learning.

[0016] A fourth aspect of the present invention provides an electronic device comprising at least one processor and a memory connected to the processor, the memory being used to store a computer program; the processor being used to execute the computer program, so that the electronic device can implement the above-mentioned sunspot counting method based on deep learning.

[0017] Compared with the existing technology, one or more of the above technical solutions have the following beneficial effects: 1. By comprehensively applying image segmentation models to segment sunspots and cluster to distinguish umbras and penumbras, and image recognition models to detect sunspot groups, the image processing process is integrated with solar physics observation standards to achieve a systematic upgrade from scientific logic to technical paths.

[0018] 2. Since sunspots always appear in active regions, we cut out the portion of the full-disk white-light image corresponding to the magnetically active regions to prevent noise points in the quiet regions from being identified as sunspots. Using image segmentation, we create pseudo-labels and train a sunspot segmentation model, replacing manual labeling and fundamentally improving labeling efficiency.

[0019] 3. During pseudo-label generation, an adaptive threshold based on the Gaussian distribution of light intensity is superior to a fixed threshold and can detect low-contrast sunspots. Pseudo-labels are corrected using historical data to filter out non-physical noise and improve reliability. The minimum sunspot area threshold is dynamically adjusted based on the cosine law of heliospheric coordinates to ensure physically reasonable results.

[0020] 4. The morphological segmentation model (MorphACWE) is introduced in conjunction with the deep learning image segmentation method to achieve segmentation effects with strong edge robustness and effective performance under weak contrast.

[0021] 5. During the counting process, the umbra is considered to exist independently, while the penumbra is not. Combined with the group bounding box training based on historical data annotation, the position and number of sunspot groups are directly output, avoiding the need to manually define group merging rules (such as distance thresholds).

[0022] 6. To address the issue of "no analytical framework that meets the Wolf number calculation standards has yet to be established", this solution not only constructs an analytical framework, but also aligns it with historical observation data to enhance the interpretability and standardization of automatic recognition results.

[0023] 7. To address the issue of "lack of physical constraints", the proposal introduces magnetic field activity area data as a reference for the physical boundaries of the active area for region of interest extraction; performs edge darkening correction on the input image to improve the segmentation effect of the perihelion disk edge area; considers the longitude and latitude of the sunspot when proposing false sunspots, and calculates the number of pixels that the false sunspots should occupy; and adopts the YOLO model for group detection, using the sunspot group classification in historical data as training labels, making it closer to the human classification standard based on spatial structure and sunspot distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0025] Figure 1 This is a schematic diagram of the overall process of a sunspot counting method based on deep learning provided by one or more embodiments of the present invention; Figure 2 is a schematic diagram of a pseudo-label generation process during sunspot counting provided by one or more embodiments of the present invention; Figure 3 is a schematic diagram of a sunspot segmentation process during sunspot counting provided by one or more embodiments of the present invention; Figure 4 is a schematic diagram of sunspot counting provided by one or more embodiments of the present invention; Figure 5 is a schematic diagram showing the correlation comparison results between the estimated total sunspot number and SIDC-SILSO and USET according to one or more embodiments of the present invention; Figure 6 is a schematic diagram of a daily comparison of the total sunspot number estimated according to one or more embodiments of the present invention and the SIDC-SILSO; Figure 7 is a schematic diagram of a daily comparison of the total sunspot number estimated and the USET according to one or more embodiments of the present invention; Figure 8 1 is a schematic diagram showing a comparison result of the correlation between the estimated total sunspot number and SWPC and DPD according to one or more embodiments of the present invention; Figure 9 is a daily comparison of the total sunspot count estimated according to one or more embodiments of the present invention and the SWPC; Figure 10 A daily comparison of the estimated total sunspot count and the DPD provided by one or more embodiments of the present invention; Figure 11 Schematic diagram of the experimental results of solar activity periodicity verification provided by one or more embodiments of the present invention (data from July 2010 to June 2023); Figure 12 This is a schematic diagram of the experimental results (test set data) of solar activity periodicity verification provided by one or more embodiments of the present invention. DETAILED DESCRIPTION

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0028] Sunspots usually appear in groups. An active area 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). Therefore, sunspot observations use the Wolf Number, which reflects the "quality" and "quantity" of solar activity through weighted calculations. It is an international standard indicator of sunspot activity.

[0029] As described in the background, traditional image processing methods (such as morphological operations and region growing algorithms) rely on manual thresholding and are insufficiently adaptable to the multi-scale structure and dynamic evolution of sunspots. While deep learning techniques based on convolutional neural networks have made progress in solar image segmentation, existing research has largely focused on single-task recognition (such as sunspot presence classification or region segmentation). An analytical framework that meets Wolf number calculation criteria has yet to be established. In particular, key aspects such as sunspot group detection and umbra / penumbra differentiation lack physical constraints.

[0030] Regarding the lack of an analytical framework that meets Wolf number standards, current sunspot counts are primarily 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 a correction factor. While most existing research focuses on a single task (such as sunspot segmentation or detection), these methods can detect the number of sunspots in images but lack a unified framework to ensure that the extracted "sunspot groups" and "individual sunspots" are consistent with the Wolf number standards defined in historical manual observations. This results in significant deviations between automatically estimated values ​​and official statistics (such as those from SIDC or NOAA).

[0031] Regarding the "lack of physical restraint mechanisms", the main points are as follows: The physical boundaries of solar active regions are unclear, and most algorithms are based on image feature segmentation, ignoring the impact of magnetic field structure on sunspot distribution. Sunspot segmentation lacks processing of optical effects such as limb darkening, which affects the identification of edge sunspots; Sunspot group classification ignores physical characteristics such as spatial continuity and life cycle.

[0032] The following examples present a deep learning-based sunspot counting method and system. They innovatively combine the MorphACWE method with a U-Net segmentation network. The segmentation results are then combined with sunspot annotation data from the Royal Observatory of Belgium (USET) database, ensuring consistency with historical ground-based observation standards. By constructing a sunspot cluster detection system for full-disk solar images based on the YOLO model and combining it with a sunspot umbra counting algorithm based on pixel intensity and K-Means, accurate counting of individual sunspots and sunspot clusters is achieved. This method effectively overcomes the limitations of limited ground-based observation data and inefficient manual annotation, providing core technical support for the construction of an automated calibration system for a cross-century sunspot database.

[0033] Example 1: This project obtains the full disk continuum spectrum observation data of the Helioseismic Magnetometer (HMI) of the Solar Dynamics Observatory (SDO) from the Joint Science Operations Center (JSOC) database.

[0034] Since SDO / HMI has been conducting routine observations since 2010, this proposal uses data from July 2010 to June 2023 as the data source for this proposal.

[0035] The data preprocessing process includes derotation correction for the north-south axial tilt of the Sun and correction for edge darkening effects. The image resolution used is 4096×4096 pixels to ensure the detection of the smallest sunspots (usually represented by only a few pixels).

[0036] To avoid potential data leakage, this solution intentionally excludes some data. The training set is constructed using observations from January 1st to September 15th of each year, while the test set consists of data from October 1st to December 15th of each year.

[0037] In order to further screen out the solar disk areas containing sunspots as the training set, some data without sunspots were excluded. The final training set contained 2464 images, and the test set covered 775 images.

[0038] In addition, this proposal uses the HMI Active Region Patch (HARP) dataset from the JSOC database for locating active regions in white-light images. This dataset provides approximate boundaries of the entire sunspot population, making it an effective tool for selecting regions of interest within full-disk observations. Both the full-disk continuum observation data and the HARP active region data are downloaded simultaneously at 12:00 UTC daily to ensure temporal consistency between the two data sets.

[0039] In addition, this solution uses manual annotations of sunspots from the USET dataset of the Royal Observatory of Belgium's Solar Equatorial Observatories to manually annotate sunspot clusters. For each SDO / HMI white-light image, the corresponding USET annotation is matched, achieving high-precision annotation of sunspot clusters by minimizing bounding boxes. These annotations are used to optimize the segmentation results of the segmentation model and serve as a training benchmark for the object detection model, thereby improving the detection accuracy of the automated framework.

[0040] like Figure 1 As shown, the sunspot counting method based on deep learning includes the following steps: Obtain image data of the solar photosphere and magnetic field active area data of the corresponding period, and pre-process them; The preprocessed image data is positioned using magnetic field activity zone data to determine the region of interest. A binary mask image is obtained through segmentation, and error correction is performed using historical data. After reconstruction, pseudo labels corresponding to the historical data are obtained. The pseudo labels obtained, together with the preprocessed image data, serve as training data for the segmentation model to obtain a reconstructed mask image. The reconstructed mask image is filtered out of the non-penumbra sunspots in the image through the pixel intensity histogram, and the non-penumbra sunspots are counted; the part with the penumbra sunspots is clustered by binary classification, and the sum of the clustered umbra number and the non-penumbra sunspot number is the number of single sunspots; The reconstructed mask image is used for image recognition based on historical data to obtain the number of sunspot groups; The obtained sunspot group number and single sunspot number are used to obtain the final sunspot counting result using the Wolf number function model.

[0041] Step 1: Pseudo label generation and model training.

[0042] S1.1 Pseudo label generation, e.g. Figure 2 As shown, this embodiment takes the HARP area of ​​NOAA No. 13186 as an example.

[0043] The original image is acquired and preprocessed, and the edge darkening correction and tilt rotation correction are performed on the SDO / HMI white light image. The coordinates of the active area are located using HARP data, and the corrected image is determined to obtain the region of interest.

[0044] MorphACWE segmentation: Based on the Gaussian distribution of light intensity in the quiet area of ​​the sun (mean μ, standard deviation σ), the intensity threshold for detecting sunspots is set to 5 times the standard deviation ( ), the number of iterations of the MorphACWE algorithm is 100, the minimum pixel number threshold is 1, and a preliminary binary mask is generated.

[0045] Historical data calibration: The sunspot group annotations produced based on the USET dataset of the Royal Observatory of Belgium were converted into binary masks. The small sunspots not defined by USET in the HARP active region were excluded by mask dot multiplication operation, and the error of the MorphACWE segmentation results was corrected.

[0046] Spatial Resolution Filtering: Considering the spatial resolution of the SDO / HMI imaging system is 0.505 arcseconds / pixel, which corresponds to approximately 362.6 km / pixel at the solar disk center, the smallest observable sunspot diameter is approximately 1500 km, which translates to approximately 4.14 pixels. To ensure the physical plausibility of the segmentation results, a minimum sunspot area threshold of 4.14 pixels is set at the solar disk center. Furthermore, the apparent size of sunspots decreases near the edge of the solar disk due to projective distortion, which follows the cosine law of heliospheric coordinates.

[0047] Specifically, the apparent pixel size of the smallest sunspot at a certain location It can be expressed as: ; in, and denote the heliospheric latitude and longitude of the location, respectively.

[0048] Therefore, if the sunspots detected by the MorphACWE algorithm have an area smaller than the minimum sunspot scale, they are considered false detections and are eliminated. After the above-mentioned spatial resolution filtering (i.e., full-disk reconstruction), the final output is a pseudo-label that is aligned with the SIDC-SILSO historical records.

[0049] The MorphACWE algorithm is an image segmentation algorithm based on morphological operations. It is an improvement on the traditional ACWE (Active Contours Without Edges) model. It includes morphological dilation and erosion steps, which replace the curvature term in traditional ACWE. It also includes regional mean calculations, similar to regional statistics in ACWE, and thresholding steps to update contours based on regional statistics. The MorphACWE algorithm achieves image segmentation by alternating these steps.

[0050] In this example, the MorphACWE algorithm is used to generate pseudo-labels and train a U-Net segmentation model. The U-Net provides the final segmentation results. Because MorphACWE excels at segmenting with blurred edges, it only requires a difference in average pixel value between the target and the background, not sharp edges. This makes it suitable for segmenting based on the morphology of sunspots in an image.

[0051] S1.2 U-Net model training, such as Figure 3 shown. Data loading: Use the PyTorch framework's DataLoader class for data loading and batch processing. Using a custom Dataset class, we encapsulate image data and pseudo-labels into a unified format and define a data loader for batch processing. The data loader's functionality includes automatic image and label loading, data normalization and standardization, and batch input to the model.

[0052] Block processing strategy: The full disk image (4096×4096) is divided into 16 1024×1024 sub-images and input into U-Net to solve the resource occupation problem and small target detection problem.

[0053] Loss function and optimization: We use binary cross entropy loss (BCELoss), Adam optimizer (learning rate 0.001), group normalization (GroupNorm) instead of batch normalization, and train for 1 epoch. Output and reconstruction: The sub-image segmentation results are filtered by spatial resolution (i.e., full disk reconstruction) and then stitched into a full disk probability map. The final segmentation mask is generated by binarization with a 0.5 threshold. This mask is multiplied element-by-element with the preprocessed image to obtain the reconstructed image.

[0054] Step 2: Count individual sunspots, such as Figure 4 (upper branch) shown.

[0055] S2.1 Umbra-penumbra distinction.

[0056] The preprocessed image is used to determine the region of interest based on the mask result. The bimodal structure is detected by the pixel intensity histogram to obtain the penumbra feature. The non-penumbra sunspots in the image are filtered out to obtain the number of non-penumbra sunspots and the part with penumbra sunspots. The part with penumbra sunspots is clustered by binary classification (umbra vs. background + penumbra). The number of clustered umbras is counted as the number of sunspots with penumbra, and added to the number of non-penumbra sunspots to obtain the number of single sunspots.

[0057] Among them, connected domain analysis: identifies connected components based on masks, detects bimodal structures (penumbra features) through pixel intensity histograms, and filters out sunspots without penumbra. Among them, K-means clustering: the retained sunspot parts are clustered into two categories (umbra vs. penumbra), and the number of clustered umbras is counted as the number of individual sunspots.

[0058] Step 3: Sunspot group detection and total number calculation, such as Figure 4 (lower branch) shown.

[0059] S3.1 YOLO sunspot detection.

[0060] Input processing: The background of the mask image is filled with the average intensity of the non-sunspot area, and the sunspot group bounding box annotated by USET is used as the label and input into the YOLOv5 model.

[0061] Model training: The image is scaled to 1024×1024, the batch size is 9, the learning rate is 0.01, and the training is carried out for 30 epochs, retaining the original loss function. Result output: Output sunspot group bounding box and count the number of detected sunspot groups.

[0062] S3.2 Total Calculation.

[0063] Formula application: Calculate the total number of sunspots according to R=S+10·G, where S (corresponding to Figure 4 Ns) is the number of single sunspots obtained in step S2.1, G (corresponding to Figure 4 Ng) in is the number of sunspot groups obtained in step S3.1.

[0064] The above solution introduces a morphological model (MorphACWE) in conjunction with U-Net to achieve a segmentation effect with strong edge robustness and effective performance under weak contrast.

[0065] The above solution can replace manual labeling: HARP + MorphACWE are used to create segmentation pseudo-labels and train the sunspot segmentation model, fundamentally improving labeling efficiency.

[0066] The above solution realizes full-process automation: it supports batch processing of full-disk sequence images, automatically outputs the number of sunspots, sunspot groups and their statistical results, and reduces dependence on manpower.

[0067] To address the problem of "no analytical framework that meets the Wolf number calculation standards has been established", this solution not only constructs an analytical framework, but also aligns it with historical observation data to enhance the interpretability and standardization of automatic recognition results.

[0068] To address the problem of "lack of physical constraint mechanism", the scheme introduces HARP data as a reference for the physical boundaries of the active area for extracting regions of interest; performs edge darkening correction on the input image to improve the segmentation effect of the perihelion disk edge area; and adopts the YOLO model for group detection, and uses the USET group division as the training label to make it closer to the human division standard based on spatial structure and sunspot distribution.

[0069] By comprehensively applying the U-Net model for sunspot segmentation, K-means clustering to distinguish umbra and penumbra, and the YOLO model to detect sunspot groups, the image processing process is integrated with solar physics observation standards to achieve a systematic upgrade from scientific logic to technical path.

[0070] In addition, a correction mask was constructed based on the Uccles Solar Equatorial Observatory (USET) dataset in Belgium to make the estimated results consistent with the historical sunspot records over the past two centuries.

[0071] Experimental results show that the daily sunspot count estimated by this method has a high correlation of 0.97 with the sunspot index provided by the Solar Impact Data Analysis Center (SIDC-SILSO), validating the robustness and reliability of this 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. It also exhibits good scalability and is suitable for future high-resolution solar observation data.

[0072] Experimental verification: Using publicly available data, we verify the consistency of this scheme with historical data, its performance across datasets, and the periodicity of solar activity.

[0073] Existing publicly available data includes: 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 ); USET: Uccle Solar Equatorial Table (USET) ( https: / / vo-tap.oma.be / ); SWPC: Space Weather Prediction Center (SWPC) (https: / / www.swpc.noaa.gov / products / solar-region-summary); DPD: Debrecen Photoheliographic Data (DPD) (http: / / fenyi.solarobs.epss.hun-ren.hu / DPD / ).

[0074] Experiment 1: Historical data consistency verification; Comparator: SIDC-SILSO, USET; Dataset: SIDC-SILSO is used as the benchmark and USET is used as the auxiliary calibration source.

[0075] Results: The estimated daily total sunspot number is consistent with the trend of SIDC-SILSO, and the Pearson correlation coefficient is 0.97 (e.g. Figure 5 and Figure 6 shown).

[0076] The estimated daily total sunspot number is consistent with the trend of USET, and the Pearson correlation coefficient is 0.93 (e.g. Figure 5 and Figure 7 shown).

[0077] Experiment 2: Cross-dataset performance comparison; Comparison objects: SWPC and DPD databases; Results: The estimated daily total sunspot number is consistent with the trend of SWPC, and the Pearson correlation coefficient is 0.94 (e.g. Figure 8 and Figure 9 shown).

[0078] The estimated daily total sunspot number is consistent with the trend of DPD, and the Pearson correlation coefficient is 0.81 (without USET mask correction) (e.g. Figure 8 and Figure 10 shown).

[0079] Experiment 3: Verification of solar activity periodicity (butterfly plot analysis); Methods: The latitude distribution of sunspot groups detected from 2010 to 2023 was statistically analyzed, and the latitude-time evolution diagram was drawn.

[0080] Results: Sunspot groups show the classic "butterfly pattern" (e.g. Figure 11 and Figure 12 That is, at the beginning of the solar cycle, sunspots are concentrated in the high-latitude region of ±30°–40°, and then gradually shift toward the equator.

[0081] Example 2: Deep learning-based sunspot counting system, including: The data acquisition and preprocessing module is configured to: acquire image data of the solar photosphere and data of the magnetic field active area in the corresponding period, and preprocess them; The pseudo label generation module is configured to: perform positioning processing on the pre-processed image data using the magnetic field activity area data to determine the region of interest, obtain a binary mask image through segmentation, perform error correction using historical data, and obtain a pseudo label corresponding to the historical data after reconstruction; The U-Net model training module is configured to use the pseudo labels and the preprocessed image data as training data for the segmentation model to obtain a reconstructed mask image. The single sunspot counting module is configured to: filter out the non-penumbra sunspots in the reconstructed mask image through the pixel intensity histogram and count the non-penumbra sunspots; perform binary clustering on the portion of the sunspots that retain the penumbra, and the sum of the clustered umbra number and the non-penumbra sunspot number 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 to obtain the number of sunspot groups; The total number calculation module is configured to: use the obtained sunspot group number and single sunspot number using the Wolf number function model to obtain the final sunspot counting result.

[0082] The MorphACWE method is combined with the U-Net segmentation network. The segmentation results are then combined with sunspot annotation data from the Royal Observatory of Belgium (USET) database to ensure consistency with historical ground-based observation standards. A sunspot cluster detection system for full-disk solar images, based on the YOLO model, is constructed, combined with a sunspot umbra counting algorithm based on pixel intensity and K-Means, to achieve accurate counting of individual sunspots and sunspot clusters. This method effectively overcomes the limitations of ground-based observation data quality and the inefficiency of manual annotation, providing core technical support for the construction of an automated calibration system for a cross-century sunspot database.

[0083] Example 3: A computer program product includes computer-readable instructions. When the computer-readable instructions are executed on an electronic device, the electronic device implements the above-mentioned sunspot counting method based on deep learning.

[0084] Example 4: An electronic device includes at least one processor and a memory connected to the processor, the memory being used to store a computer program; the processor being used to execute the computer program, enabling the electronic device to implement the above-mentioned deep learning-based sunspot counting method.

[0085] Embodiment 5: A computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enables the electronic device to implement the above-mentioned deep learning-based sunspot counting method.

[0086] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A sunspot counting method based on deep learning, characterized in that: The following steps are involved: Obtain image data of the solar photosphere and magnetic field active area data of the corresponding period, and pre-process them; The pre-processed image data is positioned using the magnetic field activity area data to determine the region of interest. After segmentation, a binary mask image is obtained. Error correction is performed using historical data, and after reconstruction, a pseudo label corresponding to the historical data is obtained. The obtained pseudo labels together with the preprocessed image data are used as training data for the segmentation model to obtain the reconstructed mask image; The reconstructed mask image is filtered out of the non-penumbra sunspots in the image through the pixel intensity histogram, and the non-penumbra sunspots are counted; the part with the penumbra sunspots is clustered by binary classification, and the sum of the clustered umbra number and the non-penumbra sunspot number is the number of single sunspots; The reconstructed mask image is used for image recognition based on historical data to obtain the number of sunspot groups; The obtained sunspot group number and single sunspot number are used to obtain the final sunspot counting result using the Wolf number function model.

2. The sunspot counting method based on deep learning as claimed in claim 1, characterized in that: A binary mask image is obtained by segmentation. Specifically, the MorphACWE segmentation algorithm is used to set the intensity threshold for detecting sunspots based on the Gaussian distribution of light intensity in the quiet area of ​​the sun to obtain a binary mask image.

3. The sunspot counting method based on deep learning as claimed in claim 1, characterized in that: Error correction is performed using historical data. Specifically, the sunspot group annotation data in the historical data is extracted and converted into a binary mask. The sunspots that do not meet the set requirements in the magnetic field active area are eliminated through mask dot multiplication operation, and the error of the MorphACWE segmentation result is corrected.

4. The sunspot counting method based on deep learning as claimed in claim 1, characterized in that: After reconstruction, the pseudo labels corresponding to the historical data are obtained. Specifically, according to the cosine law of the heliospheric coordinates and the set minimum sunspot area threshold, the apparent pixel size of the minimum sunspot at a certain position in the mask image is determined. Among the sunspots detected by the MorphACWE algorithm, if there is an area smaller than If the area is not detected, it is considered as a false detection and is removed to obtain a pseudo label corresponding to the historical data.

5. The sunspot counting method based on deep learning as claimed in claim 1, characterized in that: The reconstructed mask image determines the penumbra features through the bimodal structure in the pixel intensity histogram, and filters out the non-penumbra sunspots in the image to obtain the number of non-penumbra sunspots and the fraction of retained penumbra sunspots. For the part of the sunspots that retain the penumbra, a binary clustering of "umbra vs. background + penumbra" is performed. The number of clustered umbras is counted as the number of sunspots with penumbra, and added to the number of sunspots without penumbra to obtain the number of single sunspots.

6. The sunspot counting method based on deep learning according to claim 1, characterized in that: The reconstructed mask image is used for image recognition based on historical data to obtain the number of sunspot groups. Specifically, the background of the reconstructed mask image is filled with the average intensity of the non-spot area, and the sunspot group bounding box annotated according to the historical data is used as a label. The input is input into the YOLO model to obtain the number of sunspot groups.

7. The sunspot counting method based on deep learning according to claim 1, characterized in that: 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, where S is the number of individual sunspots obtained and G is the number of sunspot groups obtained.

8. A sunspot counting system based on deep learning, characterized by: include: The data acquisition and preprocessing module is configured to: acquire image data of the solar photosphere and data of the magnetic field active area in the corresponding period, and preprocess them; The pseudo label generation module is configured to: perform positioning processing on the pre-processed image data using the magnetic field activity area data to determine the region of interest, obtain a binary mask image through segmentation, perform error correction using historical data, and obtain a pseudo label corresponding to the historical data after reconstruction; The U-Net model training module is configured to use the pseudo labels and the preprocessed image data as training data for the segmentation model to obtain a reconstructed mask image. The single sunspot counting module is configured to: filter out the non-penumbra sunspots in the reconstructed mask image through the pixel intensity histogram and count the non-penumbra sunspots; perform binary clustering on the portion of the sunspots that retain the penumbra, and the sum of the clustered umbra number and the non-penumbra sunspot number 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 to obtain the number of sunspot groups; The total number calculation module is configured to: use the obtained sunspot group number and single sunspot number using the Wolf number function model to obtain the final sunspot counting result.

9. A computer program product, characterized in that The method comprises computer-readable instructions, which, when executed on an electronic device, enable the electronic device to implement the steps in the sunspot counting method based on deep learning as claimed in any one of claims 1 to 7.

10. An electronic device, characterized in that: The electronic device comprises at least one processor and a memory connected to the processor, wherein the memory is used to store a computer program; the processor is used to execute the computer program, so that the electronic device can implement the steps in the sunspot counting method based on deep learning as claimed in any one of claims 1 to 7.

Images