10:45 AM - 12:15 PM
[PEM09-P04] Development of image prediction method for solar active region evolution using GAN
Keywords:Space Weather, Sunspots, Deep learning
Because solar flares emit strong X-rays and plasma, causing damage to aircraft, radio systems, satellites, and so on, many studies have been conducted on flare prediction. Most of the current studies focus on the prediction of the matured active regions and the model can predict the flare occurrence only 24 hours ahead. If we can predict the growth of active regions, it would be able to predict solar flares from the predicted images at an earlier stage.
To this end, the purpose of this study is to build a growth prediction model for solar active regions using Vanilla AutoEncoder and Generative Adversarial Network, which is a deep learning model to be able to generate higher resolution images.
The active region data taken by SDO/HMI from May 2010 to April 2019 were used. For input images, we used the SHARP dataset, which is the active region magnetogram cut out from the full-disk magnetogram. Also, only active regions with sunspots large enough to be visually identified were selected to use. The dataset is composed of 194 active regions, with approximately 80% for training, 10% for validation, and 10% for testing.
We developed a GAN model that takes magnetic images as input and predicts the growth of the magnetic images after 12 hours. The difference of the total absolute flux between the predicted images and the true images on the training data was 2.33×10^21 Mx in the GAN model while it was 5.01×10^21 Mx in the AutoEncoder. The GAN model is able to generate higher resolution images compared to the AutoEncoder because MSE was reduced by about half. On the other hand, we also found a problem that the total absolute flux of the predicted images even on the training data was about 45% smaller than the total absolute flux of the true images. To solve this problem, we added the difference of the total absolute flux to the loss function. As a result, the difference between the flux values of the true and predicted images was reduced by more than 10%.
This study shows that the prediction accuracy of solar features could be improved by adding physical quantities to the loss function.
To this end, the purpose of this study is to build a growth prediction model for solar active regions using Vanilla AutoEncoder and Generative Adversarial Network, which is a deep learning model to be able to generate higher resolution images.
The active region data taken by SDO/HMI from May 2010 to April 2019 were used. For input images, we used the SHARP dataset, which is the active region magnetogram cut out from the full-disk magnetogram. Also, only active regions with sunspots large enough to be visually identified were selected to use. The dataset is composed of 194 active regions, with approximately 80% for training, 10% for validation, and 10% for testing.
We developed a GAN model that takes magnetic images as input and predicts the growth of the magnetic images after 12 hours. The difference of the total absolute flux between the predicted images and the true images on the training data was 2.33×10^21 Mx in the GAN model while it was 5.01×10^21 Mx in the AutoEncoder. The GAN model is able to generate higher resolution images compared to the AutoEncoder because MSE was reduced by about half. On the other hand, we also found a problem that the total absolute flux of the predicted images even on the training data was about 45% smaller than the total absolute flux of the true images. To solve this problem, we added the difference of the total absolute flux to the loss function. As a result, the difference between the flux values of the true and predicted images was reduced by more than 10%.
This study shows that the prediction accuracy of solar features could be improved by adding physical quantities to the loss function.