5:15 PM - 6:45 PM
[PEM11-P12] A compression method of solar polarization spectra from Hinode SOT/SP for solar flare prediction
Keywords:Solar physics, Spectro-polarimetry, Astroinformatics, Autoencoder
Forecasting solar activities, particularly the prediction of solar flares, presents both crucial and complicated challenges. The abundance of observational solar spectral data collected by the Hinode SOT/SP provides an opportunity to comprehensively explore the Sun’s behavior through advanced deep learning techniques. To address the high-dimensionality of solar polarization spectra, which poses significant processing challenges, our objective is to develop a specialized compression model for solar spectra using autoencoders. Beyond solar flare prediction, the implementation of such a model has the potential to broaden its applications including tasks, such as automatic spectra categorization and anomalous spectra detection.
We constructed both deep autoencoder (DAE) and 1D-convolutional autoencoder (CAE) models to effectively compress Stokes I and V polarization parameters. These models take polarimetric spectra as input, encode them into a compact representation, and subsequently decode them to produce an output closely resembling the input.
Our study indicates that, through experiments with various model training configurations, the CAE model exhibits greater promise in reconstructing Stokes signals, achieving a 14% reduction in deviation compared to the outcomes produced by the DAE model. We also present preliminary results of solar flare prediction using the CAE model, emphasizing its potential efficacy and discussing further improvements.
We constructed both deep autoencoder (DAE) and 1D-convolutional autoencoder (CAE) models to effectively compress Stokes I and V polarization parameters. These models take polarimetric spectra as input, encode them into a compact representation, and subsequently decode them to produce an output closely resembling the input.
Our study indicates that, through experiments with various model training configurations, the CAE model exhibits greater promise in reconstructing Stokes signals, achieving a 14% reduction in deviation compared to the outcomes produced by the DAE model. We also present preliminary results of solar flare prediction using the CAE model, emphasizing its potential efficacy and discussing further improvements.