RHESSI17--GAN

Atmospheric seeing is ubiquitous for ground-based astronomy. Atmospheric seeing produces distortions in images due to varying density and temperature structure of the Earth’s atmosphere. Bad seeing can be accounted for in part by adaptive optics built into ground-based instruments. However, with the newer generation of higher resolution ground-based instruments AO systems cannot act quickly enough to remove the worst seeing from images. As a result, we propose a generative adversarial network (GAN) which will learn how to remove blur and distortions simulated onto space-based data from SOT. The goal of this network is to generate solar flare images indistinguishable from ground-truth solar flare images. This is done through a kernel-free approach to deblurring and is a single-frame blind deconvolution method. The single-frame is important due to the low cadence of observations with respect to flare timescales meaning that a multi-frame approach could result in lost information. With the ability to generate these images, the model can then be applied to data with real seeing and they can be reconstructed with high accuracy to be included in our datasets for data analysis. The results are that spectroscopic and spectropolarimetric line profiles are successfully reconstructed by our network and so are feasible to be used for further data analysis.

大气视宁度（atmospheric seeing）是地面天文观测中普遍存在的现象。地球大气的密度与温度结构变化会使成像产生畸变，进而引发大气视宁度效应。地基天文仪器内置的自适应光学（adaptive optics）系统可在一定程度上校正较差的视宁度。然而，新一代高分辨率地基仪器的自适应光学系统响应速度不足，无法完全去除图像中最严重的视宁度干扰。为此，我们提出一种生成式对抗网络（Generative Adversarial Network，GAN），该网络可学习去除模拟在SOT空间基数据上的模糊与畸变。本网络的目标是生成与真值（ground-truth）太阳耀斑（solar flare）图像难以区分的合成太阳耀斑图像。该方法采用无核（kernel-free）盲反卷积（blind deconvolution）的单帧处理方案。鉴于太阳耀斑的时间尺度与观测时序间隔的特性，单帧处理至关重要：多帧处理方案可能会丢失观测信息。借助该图像生成能力，本模型可应用于存在真实视宁度干扰的观测数据，并以高精度完成图像重构，进而将其纳入用于数据分析的数据集。实验结果表明，本网络可成功重构光谱与光谱偏振谱线轮廓，因此可用于后续的数据分析工作。