Multi-Channel Coronal Hole Detection with Convolutional Neural Networks

Abstract

<p>Being the source region of fast solar wind streams, coronal holes are one of the key components which impact space weather. The precise detection of the coronal hole boundary is an important criterion for forecasting and solar wind modeling, but also challenges our current understanding of the magnetic structure of the Sun. We use deep-learning to provide new methods for the detection of coronal holes, based on the multi-band EUV filtergrams and LOS magnetogram from the AIA and HMI instruments onboard the Solar Dynamics Observatory. The proposed neural network is capable to simultaneously identify full-disk correlations as well as small-scale structures and efficiently combines the multi-channel information into a single detection. From the comparison with an independent manually curated test set, the model provides a more stable extraction of coronal holes than the samples considered for training. Our method operates in real-time and provides reliable coronal hole extractions throughout the solar cycle, without any additional adjustments. We further investigate the importance of the individual channels and show that our neural network can identify coronal holes solely from magnetic field data.</p>

MLMT-CNN for object detection and segmentation in multi-layer and multi-spectral images

“…Moreover, the sequential fine-tuning approach similar to [26] shows a close performance to single-band detection using faster RCNN with an identical precision, recall and F1-score over the band 304 Å and a slight decrease over the other three bands, See Table 2 and Fig. 5.…”

“…We compare against the state-of-the-art AR detector HFC's SPOCA [10]. We further compare against a sequential fine-tuning method derived from [26] through adapting the first stage of their approach to faster RCNN by sequentially fine-tuning it over the neighbouring image bands. We evaluate this approach on UAD.…”

“…This may be due to the fact that its transfer learning does not incorporate the bands' inter-dependencies when analysing the different bands. Moreover, the method was designed in [26] to produce a single prediction for the different bands, and this differs from our usage where we predict a different set of detections per band.…”

“…These observations indicate that our recursive training strategy is beneficial in cases where manual annotations are not available, such as solar ARs. 7 AR segmentation comparison between our presented method, SPOCA, and sequentially fine-tuned DNNs similar to [26], over the SPOCA subset. Red is AR, blue denotes the quite Sun background, and green is outside of the solar disk…”

“…Thus, we investigate different types of fusion and different stages to apply fusion. Another feature fusion strategy was used to segment coronal holes from SDO's 7 EUV bands and line-of-site magnetogram in [26]. The method relies on training a CNN, using weak labels, to segment coronal holes from a single band, followed by fine-tuning the learned CNN over the other bands consecutively.…”