A Machine Learning Approach to Correcting Atmospheric Seeing in Solar Flare Observations. (arXiv:2011.12814v1 [astro-ph.SR])
<a href="http://arxiv.org/find/astro-ph/1/au:+Armstrong_J/0/1/0/all/0/1">John A. Armstrong</a>, <a href="http://arxiv.org/find/astro-ph/1/au:+Fletcher_L/0/1/0/all/0/1">Lyndsay Fletcher</a>
Current post-processing techniques for the correction of atmospheric seeing
in solar observations — such as Speckle interferometry and Phase Diversity
methods — have limitations when it comes to their reconstructive capabilities
of solar flare observations. This, combined with the sporadic nature of flares
meaning observers cannot wait until seeing conditions are optimal before taking
measurements, means that many ground-based solar flare observations are marred
with bad seeing. To combat this, we propose a method for dedicated flare seeing
correction based on training a deep neural network to learn to correct
artificial seeing from flare observations taken during good seeing conditions.
This model uses transfer learning, a novel technique in solar physics, to help
learn these corrections. Transfer learning is when another network already
trained on similar data is used to influence the learning of the new network.
Once trained, the model has been applied to two flare datasets: one from
AR12157 on 2014/09/06 and one from AR12673 on 2017/09/06. The results show good
corrections to images with bad seeing with a relative error assigned to the
estimate based on the performance of the model. Further discussion takes place
of improvements to the robustness of the error on these estimates.
Current post-processing techniques for the correction of atmospheric seeing
in solar observations — such as Speckle interferometry and Phase Diversity
methods — have limitations when it comes to their reconstructive capabilities
of solar flare observations. This, combined with the sporadic nature of flares
meaning observers cannot wait until seeing conditions are optimal before taking
measurements, means that many ground-based solar flare observations are marred
with bad seeing. To combat this, we propose a method for dedicated flare seeing
correction based on training a deep neural network to learn to correct
artificial seeing from flare observations taken during good seeing conditions.
This model uses transfer learning, a novel technique in solar physics, to help
learn these corrections. Transfer learning is when another network already
trained on similar data is used to influence the learning of the new network.
Once trained, the model has been applied to two flare datasets: one from
AR12157 on 2014/09/06 and one from AR12673 on 2017/09/06. The results show good
corrections to images with bad seeing with a relative error assigned to the
estimate based on the performance of the model. Further discussion takes place
of improvements to the robustness of the error on these estimates.
http://arxiv.org/icons/sfx.gif
