10:45 〜 12:15
[PEM09-P03] A compression method of Hinode SOT/SP solar polarization data
キーワード：solar physics, specropolarimetry, high dimensional data, autoencoder
Forecasting hazardous events in the space environment is one of the main concerns in space weather research. In particular, predicting solar flare is challenging, but the task has the potential to be done by the recent computational techniques and approaches with the use of solar observational data, including solar spectra that contain important physical information to analyze.
Hinode Solar Optical Telescope-Spectropolarimeter (Hinode SOT/SP) has been accumulating solar spectro-polarimetry data for more than 15 years. However, it is difficult to process this huge amount of high dimensional (2D-space x 1D-wavelength x 1D-polarimeter) data even with the current high-speed computers. To this end, we suggest a compressed representation of SP data using a deep learning technique, and that will be useful for further steps of solar spectral analysis, such as the above-mentioned solar flare prediction, automatic categorization of spectra, detection of anomalous spectra and more.
The spectro-polarimetry data by Hinode SOT used in this study cover a wavelength range between 630.1 and 630.3 nm, including the Fe I line pair at 630.15 nm and 630.25 nm. The observation date is 2021-08-03 and the field-of-view (FoV) of the 2D spatial spectro-polarimeter is 75.6"x81.2" with a sampling slit of ~0.15". The data including both sunspots and quiet sun, high and low magnetic field regions, and were observed near the disk center (not around the solar limb) was selected.
We built an autoencoder, an encoder-decoder structured model, for compressing solar spectra containing Stokes I and V polarization parameters. The encoder converts the input of raw spectro-polarimetry data into a lower-dimensional compressed representation of the spectra, and then decodes it back into a reconstruction as the output.
We compared performances of the model trained with two different loss functions: standard loss as mean absolute error (mae), and customized loss as sum of weighted mae of Stokes I and V parameters. From the scatter plot of true and reconstructed spectral values, the model with customized loss function resulted smaller standard deviations of 0.57-0.7% (continuums) and 2.71-3.16% (line centers) for Stokes I, and 4.79% (left line core) for Stokes V.
Moreover, the error of the reconstructed spectra was larger around sunspots (active regions of the solar surface having strong magnetic fields). Assuming that the imbalance of the pixel count between the sunspot and surrounding region may cause this problem, we prepared datasets with different degrees of balance as increasing the number of pixels at active regions relative to the number of pixels at quiet regions in each dataset. The result showed that the tendency of higher the data balance, better the model performance.
Hinode Solar Optical Telescope-Spectropolarimeter (Hinode SOT/SP) has been accumulating solar spectro-polarimetry data for more than 15 years. However, it is difficult to process this huge amount of high dimensional (2D-space x 1D-wavelength x 1D-polarimeter) data even with the current high-speed computers. To this end, we suggest a compressed representation of SP data using a deep learning technique, and that will be useful for further steps of solar spectral analysis, such as the above-mentioned solar flare prediction, automatic categorization of spectra, detection of anomalous spectra and more.
The spectro-polarimetry data by Hinode SOT used in this study cover a wavelength range between 630.1 and 630.3 nm, including the Fe I line pair at 630.15 nm and 630.25 nm. The observation date is 2021-08-03 and the field-of-view (FoV) of the 2D spatial spectro-polarimeter is 75.6"x81.2" with a sampling slit of ~0.15". The data including both sunspots and quiet sun, high and low magnetic field regions, and were observed near the disk center (not around the solar limb) was selected.
We built an autoencoder, an encoder-decoder structured model, for compressing solar spectra containing Stokes I and V polarization parameters. The encoder converts the input of raw spectro-polarimetry data into a lower-dimensional compressed representation of the spectra, and then decodes it back into a reconstruction as the output.
We compared performances of the model trained with two different loss functions: standard loss as mean absolute error (mae), and customized loss as sum of weighted mae of Stokes I and V parameters. From the scatter plot of true and reconstructed spectral values, the model with customized loss function resulted smaller standard deviations of 0.57-0.7% (continuums) and 2.71-3.16% (line centers) for Stokes I, and 4.79% (left line core) for Stokes V.
Moreover, the error of the reconstructed spectra was larger around sunspots (active regions of the solar surface having strong magnetic fields). Assuming that the imbalance of the pixel count between the sunspot and surrounding region may cause this problem, we prepared datasets with different degrees of balance as increasing the number of pixels at active regions relative to the number of pixels at quiet regions in each dataset. The result showed that the tendency of higher the data balance, better the model performance.