Abstract
The topology and dynamics of the solar chromosphere are greatly affected by the presence of magnetic fields. The magnetic field can be inferred by analyzing polarimetric observations of spectral lines. Polarimetric signals induced by chromospheric magnetic fields are, however, particularly weak, and in most cases very close to the detection limit of current instrumentation. Because of this, there are only few observational studies that have successfully reconstructed the three components of the magnetic field vector in the chromosphere. Traditionally, the signal-to-noise ratio of observations has been improved by performing time-averages or spatial averages, but in both cases, some information is lost. More advanced techniques, like principal-component-analysis, have also been employed to take advantage of the sparsity of the observations in the spectral direction. In the present study, we propose to use the spatial coherence of the observations to reduce the noise using deep-learning techniques. We design a neural network that is capable of recovering weak signals under a complex noise corruption (including instrumental artifacts and non-linear post-processing). The training of the network is carried out without a priori knowledge of the clean signals, or an explicit statistical characterization of the noise or other corruption. We only use the same observations as our generative model. The performance of this method is demonstrated on both, synthetic experiments and real data. We show examples of the improvement in typical signals obtained in current telescopes such as the Swedish 1-meter Solar Telescope. The presented method can recover weak signals equally well no matter on what spectral line or spectral sampling is used. It is especially suitable for cases when the wavelength sampling is scarce.
