11:00 AM - 1:00 PM
[PEM09-P12] Development of a data compression method for Sunspot magnetograms using Autoencoder
Keywords:Space weather, Deep learning, Big Data
Recent evoltion of machine learning enables us to predict solar flare occurrence with a highly accuracy. The CNN model achieves higher accuracy than the MLP model and this can be interpreted that the neural networks are able to distinct the features from the magnetogram better than human knowledge so far. Further, it is reported that the accuracy of the flare prediction becomes higher with the input of the time series information. However, only the table-based information, such as the history of flare occurrence, is used so far. As a next step, it is natural to use time series of image data as the input, but this is not practical now because it requires a huge amount of computation. One of the solutions to this problem is to compress the input image. To this end, we try constructing a data compression method of solar magnetograms by Autoencoder and develop a solar flare prediction model with an enough accuracy.
We used sunspot magnetograms taken by SDO/HMI from May 2010 to April 2011. The SHARP dataset, the active region magnetogram cut out from the full-Sun magnetogram, was used the input image and the SWAN dataset was used as the flare label. We define that the flare “occurs“ in this study when at least one flare occurs in the magnetogram within 24 hours. We reduced the data cadence to 1 hour and used 90% for training and 10% for testing. In addition, data shuffling is used between the training and testing.
Our model construction consists of two steps. In the first step, we develop the Convolutional Autoencoder model for Sunspot magnetograms by reproducing the input image. In the second step, the flare prediction model is developed with an input of the compressed feature vector produced by the Autoencoder. The developed model achieved TSS = 0.730 for solar flare prediction above C class and the input data volume is compressed by a factor of two. The highest score in the previous studies is TSS=0.724 with uncompressed data, and our results succeeded to achieve a similar score with compressed data.
We used sunspot magnetograms taken by SDO/HMI from May 2010 to April 2011. The SHARP dataset, the active region magnetogram cut out from the full-Sun magnetogram, was used the input image and the SWAN dataset was used as the flare label. We define that the flare “occurs“ in this study when at least one flare occurs in the magnetogram within 24 hours. We reduced the data cadence to 1 hour and used 90% for training and 10% for testing. In addition, data shuffling is used between the training and testing.
Our model construction consists of two steps. In the first step, we develop the Convolutional Autoencoder model for Sunspot magnetograms by reproducing the input image. In the second step, the flare prediction model is developed with an input of the compressed feature vector produced by the Autoencoder. The developed model achieved TSS = 0.730 for solar flare prediction above C class and the input data volume is compressed by a factor of two. The highest score in the previous studies is TSS=0.724 with uncompressed data, and our results succeeded to achieve a similar score with compressed data.