A solar burst prediction method based on multi-modal solar image and physical feature fusion
Patent Information
- Application Number
- CN202611099801.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-23
- Publication Date
- 2026-10-27
AI Technical Summary
[0005]针对现有技术，本发明提供一种基于多模态太阳图像和物理特征融合的太阳爆发预测方法，以解决多源太阳数据统一建模困难、活动区关键前兆稀疏、单纯图像模型可解释性不足以及全日面与活动区预报割裂的问题
[0025]本发明预测方法中，通过将多源太阳观测数据统一为具有明确物理含义的三通道输入；同时支持全日面和活动区两个尺度的太阳爆发预测；通过物理特征和支持向量机判别增强深度视觉模型的可解释性；输出结果可直接接入空间天气业务平台。
Smart Images
Figure CN122990399A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a combination of space weather forecasting, solar image processing, artificial intelligence and machine learning technologies, and in particular to a method for predicting solar flares based on the fusion of multimodal solar images and physical features. Background Technology
[0002] Solar flares, coronal mass ejections, and high-energy solar particle events are significant sources of space weather disturbances, affecting satellite operations, radio communications, power systems, and manned space missions. Predicting solar flares requires comprehensive analysis of multi-source information, including photosphere magnetic fields, coronal radiation, historical eruptive activity, and physical parameters of active regions.
[0003] Existing solar eruption prediction methods typically rely on a single image modality or a small number of manually designed parameters, making it difficult to simultaneously characterize both long-range spatial correlations across the entire solar surface and local eruption precursors in active regions. While deep learning models can handle complex image structures, the operational interpretability and stability of prediction results are insufficient without the fusion of physical parameters.
[0004] Furthermore, solar eruption samples are relatively sparse, and different observational data vary in spatial resolution, physical meaning, noise distribution, and timestamp organization, making it difficult to establish a unified engineering workflow for full solar surface prediction and active region prediction. Therefore, it is necessary to propose a solar eruption prediction scheme that combines multimodal solar image representation, active region localization, and physical feature discrimination. Summary of the Invention
[0005] In response to existing technologies, this invention provides a solar eruption prediction method based on the fusion of multimodal solar images and physical features, in order to solve the problems of difficulty in unified modeling of multi-source solar data, sparse key precursors of active regions, insufficient interpretability of simple image models, and the disconnect between the forecast of the entire solar surface and the forecast of active regions.
[0006] To address the aforementioned technical problems, this invention proposes a solar eruption prediction method based on the fusion of multimodal solar images and physical features, mainly comprising:
[0007] S1. Acquire solar observation data at the same timestamp, the solar observation data including magnetograms acquired by the Helioseismic and Magnetic Imager (HMI), 304 angstrom images and flare index masks acquired by the Atmospheric Imaging Assembly (AIA);
[0008] S2. Perform timestamp registration and spatial scale unification on the HMI magnetograph, AIA 304 angstrom image and flare index mask, and map the three types of data to different channels of the three-channel image to generate a full-surface multimodal solar image;
[0009] S3. Input the full-surface multimodal solar image into the visual coding model to obtain the full-surface solar eruption visual prediction probability;
[0010] S4. Read the active region number mask, generate an active region cropping image for the valid active region number, and input the active region cropping image into the visual coding model to obtain the visual prediction probability of active region-level solar flares;
[0011] S5. For full solar surface prediction or active area-level prediction, match solar physical features by date or by date and active area number. The solar physical features include at least one of the following: 10.7 cm solar radio current (F10.7), R value characterizing the weighted magnetic flux of the high gradient magnetic field polarity reversal line, flare index statistics, and active area.
[0012] S6. Input the solar physical characteristics into the machine learning discrimination model to obtain the physical discrimination value;
[0013] S7. Perform weighted fusion based on the visual prediction probability and the physical discrimination value to obtain a fusion prediction score;
[0014] S8. Output the solar eruption prediction results at the scale of the entire solar surface and / or active region within the target prediction period based on the fusion prediction score.
[0015] Furthermore, in the solar eruption prediction method of the present invention, the flare index (FI) mask is constructed only based on historical flare events prior to the prediction reference time. The flare index (FI) is calculated according to FI = N_C + 10N_M + 100N_X, where N_C, N_M, and N_X are the number of Class C, Class M, and Class X flares occurring in the corresponding active region within a preset historical period, respectively. The target prediction period is 24 hours after the prediction reference time, and the solar eruption is defined as a Class M or higher solar flare occurring within the target prediction period.
[0016] Furthermore, in the solar eruption prediction method of the present invention, wherein,
[0017] In step S2, the flare index mask is unified to the same spatial resolution as the HMI magnetograph and the AIA 304 angstrom image before generating the three-channel image, and is truncated and normalized according to a preset intensity range; the red channel of the three-channel image corresponds to the HMI magnetograph, the green channel corresponds to the AIA 304 angstrom image, and the blue channel corresponds to the flare index mask.
[0018] In step S3, the visual coding model includes an image patch input adapter, a Transformer encoder, and a linear output adapter; the image patch input adapter divides the input image into multiple image patches and generates image labels; the Transformer encoder extracts spatial features; and the linear output adapter outputs binary classification log odds output values and normalized probabilities.
[0019] In step S4, after reading the active region number mask, background numbers and invalid numbers are excluded. A bounding box is calculated for each valid active region number. The corresponding region is cropped from the full-plane multimodal solar image and generated as an active region cropped image after background filling and size normalization.
[0020] In step S5, for the whole solar surface prediction, the matched solar physical features include the matching F10.7 solar radio current, the maximum value of R value in each active region on the day, and the maximum pixel value in the flare index mask; for the active region level prediction, the matched solar physical features include the matching F10.7 solar radio current, the R value of the corresponding active region, the local flare index, and the area of the active region.
[0021] In step S6, the machine learning discrimination model is a support vector machine (SVM) model; the missing value imputer and normalizer obtained by fitting the training samples are used to process the solar physical features, and the physical discrimination value is obtained through the support vector machine decision function.
[0022] The weighted fusion in step S7 is calculated according to F = αD + (1 - α)(P - 0.5), where F is the fused prediction score, D is the decision function value of the support vector machine after direction unification, P is the visual prediction probability of solar eruption output by the visual coding model, and α is the fusion weight. This invention does not limit the specific value of the fusion weight. It is suggested that the value of α be in the range of 0 to 1.
[0023] In step S8, the fused prediction score is compared with a preset decision threshold, and the prediction category is output; label 0 represents the eruption category, and label 1 represents the non-eruption category. The solar eruption prediction results output in step S8 include sample identifier, visual prediction probability, physical discrimination value, fused prediction score, and prediction category, and are output in the form of data file, database record, or business interface response.
[0024] Compared with the prior art, the beneficial effects of the present invention are:
[0025] In the prediction method of this invention, multi-source solar observation data is unified into three-channel input with clear physical meaning; it supports solar eruption prediction at two scales: the entire solar disk and the active region; the interpretability of the deep visual model is enhanced by physical features and support vector machine discrimination; and the output results can be directly connected to the space weather service platform. Attached Figure Description
[0026] Figure 1 This is a flowchart of the steps of the solar eruption prediction method described in this invention;
[0027] Figure 2 This is a flowchart illustrating the overall process of predicting solar flares according to an embodiment of the present invention.
[0028] Figure 3 This is an active region processing and physical feature matching diagram in the solar eruption prediction method of this invention.
[0029] Figure 4 This is a diagram illustrating the multimodal solar image visual coding structure in the solar eruption prediction method of this invention.
[0030] Figure 5 This is a flowchart of the visual probability and physical feature fusion prediction process in the solar eruption prediction method of this invention.
[0031] Figure 6 This is a flowchart illustrating the application of the solar eruption prediction method in an embodiment of the present invention.
[0032] Figure 7 The image shows the experimental results of the multimodal solar flare prediction model according to an embodiment of the present invention. Detailed Implementation
[0033] This invention provides a solar eruption prediction method based on the fusion of multimodal solar images and physical features. The method mainly includes: acquiring HMI magnetographs, AIA 304 angstrom images, and flare index masks at the same timestamp; mapping the three types of observation data into three-channel images after registration and normalization; outputting the visual prediction probability of solar eruptions using a visual coding model; generating active region cropped images based on active region number masks and performing active region-level visual prediction; matching physical features such as F10.7 solar radio flux, R-value, flare index, and active region area by date or active region number; obtaining physical discriminant values through a support vector machine model; and weightedly fusing the visual prediction probability and the physical discriminant values to output the solar eruption prediction result. Figure 1 As shown, the data generated by S1 to S8 are sequentially passed to the next processing step, as detailed below:
[0034] S1. Acquire solar observation data at the same timestamp, including HMI magnetograph, AIA 304 angstrom image and flare index mask.
[0035] S2. Perform timestamp registration and spatial scale unification on the HMI magnetograph, AIA 304 angstrom image and flare index mask, and map the three types of data to different channels of the three-channel image to generate a full-surface multimodal solar image;
[0036] The flare index mask uses only historical flare events prior to the prediction baseline time and is unified to the same spatial resolution as the HMI magnetograph and the AIA 304 Å image before generating the three-channel image. The red, green, and blue channels of the three-channel image correspond to the HMI magnetograph, the AIA 304 Å image, and the flare index mask, respectively, thereby unifying the magnetic field structure, extreme ultraviolet radiation morphology, and historical eruption clues into the same image tensor.
[0037] S3. Input the full-surface multimodal solar image into the visual coding model to obtain the full-surface solar eruption visual prediction probability; the visual coding model includes an image patch input adapter, a Transformer encoder and a linear output adapter, which sequentially complete image tag generation, spatial feature extraction and binary classification log probability output value and normalized probability output.
[0038] S4. Read the active region number mask, exclude background numbers and invalid numbers, calculate the bounding box for each valid active region number, crop the corresponding region from the full-plane multimodal solar image, and generate an active region cropped image after background filling and size normalization; input the active region cropped image into the visual coding model to obtain the visual prediction probability of active region-level solar flares.
[0039] S5. For full solar surface prediction or active region-level prediction, match solar physical features by date or by date and active region number. Full solar surface physical features include F10.7 solar radio current, the maximum value of R values among all active regions on that day, and the maximum pixel value in the flare index mask; active region-level physical features include F10.7 solar radio current, the R value of the corresponding active region, the local flare index, and the area of the active region.
[0040] S6. Perform missing value imputation and standardization on the solar physical features and input them into the support vector machine model. Obtain the physical discriminant value through the support vector machine decision function. The missing value imputer and standardizer are fitted only by the training samples.
[0041] S7. Perform weighted fusion based on the visual prediction probability and the physical discrimination value after direction unification, and calculate the fusion prediction score according to F = αD + (1 - α)(P - 0.5), where F, D, P and α represent the fusion prediction score, the support vector machine decision function value, the visual prediction probability of the solar eruption and the fusion weight, respectively; the present invention does not limit the specific value of the fusion weight.
[0042] S8. Output the solar eruption prediction results based on the fusion prediction score; label 0 indicates the eruption category, and label 1 indicates the non-eruption category. The prediction results include sample identifier, visual prediction probability, physical discrimination value, fusion prediction score, and prediction category, and are output in the form of data file, database record, or business interface response.
[0043] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the following embodiments are by no means intended to limit the present invention.
[0044] Example: The solar eruption prediction method provided in this example can be executed according to a date or a preset task cycle, and performs registration, three-channel sample construction, active region clipping, visual encoding, physical feature discrimination and fusion output on the input solar observation data. Figure 2 The overall flow of the solar eruption prediction method in this embodiment is illustrated. After timestamp registration, spatial scale unification, and three-channel image construction, the HMI magnetograph, AIA 304 Å image, and flare index mask are input into the full-surface image visual coding prediction branch and the active region cropped image visual coding prediction branch, respectively. The physical feature branch receives F10.7 solar radio flux, R-value, flare index, and active region area, and sequentially performs missing value imputation, standardization, and support vector machine discrimination to obtain physical discrimination values. The visual prediction probabilities generated by the two visual branches and the physical discrimination values are combined in a weighted fusion step to form a prediction category, risk score, and corresponding explanatory features, and the prediction result is output. The outputs of each step serve as inputs for the next step, forming a continuous data transfer relationship. The specific process is explained in detail below:
[0045] Data access steps: The following data were obtained from publicly available observational data products from the Solar Dynamics Observatory, soft X-ray flare event data, daily solar radio flux files, and active region identification results: HMI magnetograph, AIA 304 angstrom image, flare index mask, active region number mask, F10.7 solar radio flux, R-value, and active region parameters. Data was organized using the timestamp yyyyMMdd_HHmmss, where yyyy, MM, dd, HH, mm, and ss represent year, month, day, hour, minute, and second, respectively.
[0046] The time stamp registration and three-channel image generation steps are as follows: HMI magnetograph, AIA 304 Å image and flare index mask with the same time stamp are matched, the flare index mask is unified to the preset spatial resolution, and the three types of data are normalized; then the three types of data are written into the red, green and blue channels respectively to generate a full-surface multimodal solar image.
[0047] Cutting steps for the activity area: as follows Figure 3As shown, the active region number mask is read, background and invalid numbers are filtered out, and the bounding box and region of interest (ROI) clipping range are calculated for each valid active region number. The clipped image is then filled with background and normalized to form an active region image of a preset size. Based on the same date and active region number, the F10.7 solar radio flux, R-value, flare index, and active region area are matched to form a physical feature vector. After missing value imputation and standardization, the physical feature vector is input into a support vector machine to obtain physical discriminant values, which are then fused with the visual prediction results of the active region to form an active region-level prediction result.
[0048] Visual encoding prediction steps: such as Figure 4 As shown, the input is a three-channel solar image, where the red channel carries the HMI magnetograph, the green channel carries the AIA 304 image, and the blue channel carries the flare index mask. The image patch input adapter divides the image into a sequence of image patches of a preset input size and generates image labels; the Transformer encoder extracts global spatial features through multi-layer self-attention encoding; and the linear output adapter outputs a binary log-odds output value and a visual prediction probability of solar eruptions. The visual prediction probability characterizes the solar eruption risk corresponding to the input image and serves as the visual branch input for the weighted fusion decision.
[0049] Physical feature extraction steps: For full-area prediction, match the F10.7 solar radio flux, the maximum R-value among active regions for the day, and the maximum pixel value in the flare index mask by date; for active region-level prediction, match the F10.7 solar radio flux, R-value, local flare index, and active region area by date and active region number. Missing features are handled by a missing value imputer fitted only with training samples and standardized by a normalizer.
[0050] Feature preprocessing and support vector machine discrimination steps: as follows Figure 5 As shown, multimodal solar images of the entire solar disk or active region are input into the visual encoding model to obtain visual prediction probabilities. The visual branch uses the value of the solar eruption visual prediction probability minus 0.5. The F10.7 solar radio flux, R value, flare index, and active region area are input into the support vector machine model after missing value imputation and standardization to obtain directionally unified physical discriminant values. The visual branch input and the physical discriminant values are weighted and fused to form a fusion prediction score and a fusion prediction decision, outputting the prediction category, risk score, and explanatory features; in the figure, the explanatory features are the physical feature values participating in the current sample's discrimination.
[0051] Prediction output steps: Output the predicted category based on the fusion prediction score, and generate a result record containing sample identifier, visual prediction probability, physical discrimination value, fusion prediction score and prediction category; the result record is transmitted to subsequent business steps in the form of data file, database record or business interface response.
[0052] In the daily prediction embodiment of strong solar flares across the entire solar surface, the method generates a three-channel image of the entire solar surface according to the timestamp, obtains the visual prediction probability of solar flares through a visual coding model, and inputs the physical features of the entire solar surface into a support vector machine model to obtain physical discrimination values. The two are then weighted and fused to output the daily prediction result of solar flares across the entire solar surface.
[0053] In the active region-level strong flare prediction embodiment, the method generates an active region cropped image based on the active region number mask, obtains the visual prediction probability of the active region through a visual coding model, and simultaneously matches physical features such as R value, local flare index and active region area according to date and active region number. After discrimination by support vector machine and weighted fusion, the solar eruption prediction results of each active region are output.
[0054] In the integrated implementation of the business forecasting platform, such as Figure 6 As shown, after solar observation data is accessed periodically, the prediction method executes according to a predetermined cycle and generates result records. These records include total solar risk, active area risk, fused prediction score, prediction category, and corresponding physical interpretation characteristics. Forecasters review and assess the risk based on the risk score and labels, and the reviewed results are output through the early warning release interface.
[0055] Figure 7 The experimental results of the multimodal solar flare prediction model are shown. In the model validation experiment, the period from January 2023 to May 2024 was used as an independent test period, and the prediction target was whether a solar flare of class M or above would occur in the next 24 hours. The area under the receiver operating characteristic curve (ROC-AUC) was used to evaluate the discriminative ability of the probabilistic forecast: at the activity region scale, the ROC-AUC of the present invention model and the Space Weather Prediction Center (SWPC) were 0.8368 and 0.8145, respectively; at the all-solar scale, the ROC-AUC of the present invention model, the Space Environment Prediction Center (SEPC) of the Chinese Academy of Sciences, and SWPC were 0.7407, 0.7301, and 0.8014, respectively. The binary classification evaluation uses F1 score (F1), True Skill Statistic (TSS), and accuracy. The corresponding values for the active area scale are 0.4384, 0.5009, and 0.9001, respectively, and the corresponding values for the full-day scale are 0.6563, 0.4430, and 0.7417, respectively.
[0056] To illustrate the roles of each technical branch, a visual model comparison method and a physical feature comparison method can be set up: the visual model comparison method only uses multimodal solar images to output prediction probabilities, while the physical feature comparison method only uses solar physical features for discrimination. This invention simultaneously utilizes image spatial structure, historical flare activity, and quantitative physical features to form prediction results at two scales: the entire solar disk and the active region.
[0057] The method of this invention is not limited to a specific execution environment, as long as it can perform image preprocessing, visual encoding model inference, physical feature preprocessing, support vector machine discrimination, and fusion decision-making. The visual encoding model, support vector machine model, missing value imputer, and normalizer can be periodically updated according to newly added observation data.
[0058] Although the present invention has been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are preferred application examples that demonstrate the core technical ideas of the present invention, and are merely illustrative and not restrictive. Equivalent substitutions or modifications made by those skilled in the art to the data source, model structure, fusion weights, output interface and deployment method without departing from the spirit of the present invention are all within the protection scope of the present invention.
Claims
1. A method for predicting solar eruptions based on the fusion of multimodal solar images and physical features, characterized in that, Includes the following steps: S1. Acquire solar observation data at the same timestamp, the solar observation data including HMI magnetograph acquired by the helioseismic and magnetic imager, AIA 304 angstrom image and flare index mask acquired by the atmospheric imaging component; S2. Perform timestamp registration and spatial scale unification on the HMI magnetograph, AIA 304 angstrom image and flare index mask, and map the three types of data to different channels of the three-channel image to generate a full-surface multimodal solar image; S3. Input the full-surface multimodal solar image into the visual coding model to obtain the full-surface solar eruption visual prediction probability; S4. Read the active region number mask, generate an active region cropping image for the valid active region number, and input the active region cropping image into the visual coding model to obtain the visual prediction probability of active region-level solar flares; S5. For full solar surface prediction or active area prediction, match solar physical features by date or by date and active area number. The solar physical features include at least one of the following: 10.7 cm solar radio flux, R value representing the weighted magnetic flux of the high gradient magnetic field polarity reversal line, flare index statistics, and active area. S6. Input the solar physical characteristics into the machine learning discrimination model to obtain the physical discrimination value; S7. Perform weighted fusion based on the visual prediction probability and the physical discrimination value to obtain a fusion prediction score; S8. Output the solar eruption prediction results at the scale of the entire solar surface and / or active region within the target prediction period based on the fusion prediction score.
2. The solar eruption prediction method according to claim 1, characterized in that, The target prediction period is 24 hours after the prediction baseline time, and the solar eruption is an M-class or higher solar flare occurring within the target prediction period. The flare index mask is constructed based on historical flare events prior to the prediction baseline time, and does not use flare events within the target prediction period; the flare index is denoted as FI, and is calculated according to FI = N_C + 10N_M + 100N_X, where N_C, N_M, and N_X are the number of Class C, Class M, and Class X flares occurring in the corresponding active area within the preset historical period, respectively.
3. The solar eruption prediction method according to claim 1 or 2, characterized in that, In step S2, the flare index mask is unified to the same spatial resolution as the HMI magnetograph and the AIA 304 angstrom image before generating the three-channel image, and is truncated and normalized according to a preset intensity range; the red channel of the three-channel image corresponds to the HMI magnetograph, the green channel corresponds to the AIA 304 angstrom image, and the blue channel corresponds to the flare index mask.
4. The solar eruption prediction method according to claim 1, characterized in that, In step S3, the visual coding model includes an image patch input adapter, a Transformer encoder, and a linear output adapter; the image patch input adapter divides the input image into multiple image patches and generates image labels; the Transformer encoder extracts spatial features; and the linear output adapter outputs binary classification log odds output values and normalized probabilities.
5. The solar eruption prediction method according to claim 1, characterized in that, In step S4, after reading the active region number mask, background numbers and invalid numbers are excluded. A bounding box is calculated for each valid active region number. The corresponding region is cropped from the full-plane multimodal solar image and generated as an active region cropped image after background filling and size normalization.
6. The solar eruption prediction method according to claim 1, characterized in that, In step S5, the 10.7 cm solar radio current is denoted as F10.7 solar radio current. For full solar projection, the matched solar physical features include F10.7 solar radio flux, the maximum value of R values in each active region on that day, and the maximum pixel value in the flare index mask. For active region-level predictions, the matched solar physical features include F10.7 solar radio flux, the R value of the corresponding active region, the local flare index, and the area of the active region.
7. The solar eruption prediction method according to claim 1, characterized in that, In step S6, the machine learning discrimination model is a support vector machine model; the missing value imputer and normalizer obtained by fitting the training samples are used to process the solar physical features, and then the physical discrimination value is obtained through the support vector machine decision function.
8. The solar eruption prediction method according to claim 7, characterized in that, The weighted fusion in step S7 is calculated according to F = αD + (1 - α)(P - 0.5), where F is the fusion prediction score, D is the support vector machine decision function value after direction unification, P is the visual prediction probability of solar eruption output by the visual coding model, and α is the fusion weight with a value ranging from 0 to 1.
9. The solar eruption prediction method according to claim 1, characterized in that, The solar eruption prediction results output in step S8 include sample identifiers, visual prediction probabilities, physical discrimination values, fusion prediction scores, and prediction categories, and are output in the form of data files, database records, or business interface responses.