125. Solar Farside Magnetograms from Deep Learning Analysis of STEREO/EUVI Data

Contributed by Yong-Jae Moon. Posted on April 29, 2019

Yong-Jae Moon, Taeyoung Kim, Eunsu Park, and Harim Lee
School of Space Research, Kyung Hee University

Solar magnetograms are important for studying solar activity and predicting space weather
disturbances. Farside active regions can be constructed using local helioseismology without any farside data1, and their quality is lower than typical frontside magnetograms.

We apply a deep learning model based on conditional generative adversarial networks2 to solar magnetograms and EUV images. For training and evaluation data sets, we consider pairs of SDO/AIA 304 Å images and SDO/HMI line-of-sight magnetograms with 12 hours cadence from 2011 to 2017. As a result, we have 4,972 pairs of SDO/AIA images and SDO/HMI magnetograms. We select 4,147 pairs from 2011 to 2017 except for Septembers and Octobers for the training data set and 825 pairs in Septembers and Octobers for the evaluation data set.

Next we evaluate how well our model generates magnetograms. Figure 1 shows AIA 304 Å images used as the input data, AI-generated magnetograms, and HMI magnetograms. A comparison shows that the bipolar structures of the HMI magnetograms were well restored. The quality of the AI-generated magnetograms is highly comparable with that of the HMI ones. Even though we do not have any priori condition, bipolar structures in AI-generated magnetograms mostly follow Hale’s law, which is an observational rule: one polarity preceding the other polarity in the northern hemisphere and vice versa in the southern hemisphere. In the training step, the generator is trained to learn the polarity patterns of active regions. In the evaluation and generation step, the generator reproduces the pattern. Since all data are from the Solar Cycle 24, there is no problem producing the Hale’s law pattern in this cycle. Note that the polarity of the solar magnetic field is reversed cycle by cycle, hence, our model that is trained for Cycle 24 would only be effective for even solar cycles, but should be tested for odd cycles. A careful comparison between two magnetograms shows that the tilt angle between one preceding sunspot and the following one is not always properly generated, which is a limitation of our method. The EUV 304 Å emission is from the chromosphere-transition region, whereas the HMI measures the magnetic field in the photosphere. The discrepancy in the detailed structures between the real magnetogram and AI-generated magnetogram reflects the difficulty of the present model in precisely reconstructing the photospheric magnetic fields from chromosphere emission signatures.

Figure 1| Comparison between HMI magnetograms and AI-generated ones from SDO/AIA 304 Å images. a: the AIA images, which are input data of the generator, from September 1 to 7, 2017 with 2 days cadence, b: the AI-generated magnetograms from AIA images using the model, and c: the HMI images as the ground truth.

 

Then we apply our model to EUVI 304 Å images onboard STEREO, whose filter response function is consistent with that of AIA images, in order to generate farside magnetograms. In the case of STEREO-B, data are only available before October 1, 2014, due to multiple hardware anomalies affecting the control of the spacecraft orientation. Figure 2 shows a series of EUV 304 Å images by STEREO-B/EUVI and AIA, and magnetograms from June 4 to June 13, 2014: two farside magnetograms generated by our model and two HMI ones. On June 4, the STEREO-B is located on -164 heliographic longitudinal degrees from the central meridian, which makes STEREO-B images mostly farside ones. It is evident that the parallel bipolar structures in NOAA AR 12087, as indicated by the yellow box in Figure 2, are well generated and conserved during the time of observations. Thus, we can successfully monitor the temporal evolution of this active region from the farside to the frontside when farside EUV data are available.

Figure 2| A series of 304 Å images (top) and magnetograms (bottom).Top: the first two 304 Å images are taken from STEREO-B/EUVI and the last two from SDO/AIA. Bottom: First two magnetograms are AI-generated farside ones from the model and the last two are taken from SDO/HMI. The yellow boxes show the tracking of the NOAA AR 12087 from the farside to the frontside.

It is interesting to monitor the continuous evolution of active regions from the farside to the frontside. Figure 3 shows the temporal evolution of total unsigned magnetic flux of NOAA active region 12087 from June 3 to June 19 2014. As seen in the figure, the total magnetic flux of this active region taken from the farside is about two times larger than that from the frontside. It is noted that the measurements of magnetic fluxes near the eastern limb are significantly underestimated because of the projection effects. The temporal evolution is very valuable from view of space weather forecast in that there were three consecutive strong flares (X 2.2, 1.5, and 1.0) near the limb on June 10-11, 2014. We estimate the uncertainty of magnetic flux from farside magnetograms. In fact, it is not easy to conduct an inter-calibration between AIA and EUVI since they could not observe the same field of view.

Figure 3| Temporal evolution of total unsigned magnetic flux of the NOAA AR 12087 from June 3 to 19 2014. Filled diamonds represent the total unsigned magnetic flux from AI-generated magnetograms using STEREO-B/EUVI 304 Å images. Error bars indicate their relative errors. Downward arrows indicate that AR12087 located at 60 and 100 heliographic longitudinal degrees east of the central meridian, respectively. Filled circles represent the total unsigned magnetic flux from SDO/HMI magnetograms. Solid curve represents 5-min averaged GOES-15 X-ray flux (0.1-0.8 nm).

In this study, the proposed model shows sufficient potential for image-to-image translation between two different scientific sensor images. In astronomy and geophysical fields, many multi-wavelength observations have been made, so the model can be used to extend these kinds of data. This methodology can also be applied to a variety of scientific fields that use different kinds of sensor images.

Please refer to Ref. [3] for details.

References

[1] Lindsey, C. & Braun, D. C., 2000, Science, 287, 1799
[2] Isola, P., Zhu, J.-Y., Zhou, T., et al. 2017, CVPR IEEE, 2017, 1125
[3] Kim, T., Park, E., Lee, H., et al. 2019, Nature Astronomy, DOI: 10.1038/s41550-019-0711-5

Leave a comment

Your email address will not be published. Required fields are marked *