Automatic detection of Ellerman bombs using Deep Learning

TL;DR

This study develops neural network models for automatic detection of Ellerman bombs in solar observations, highlighting the importance of spectral and spatial resolution, and the potential of temporal data for improved detection in large datasets.

Contribution

The paper introduces neural network-based methods for detecting Ellerman bombs in solar data, analyzing the impact of spectral, spatial, and temporal information on detection performance.

Findings

NN models effectively detect EBs in high-resolution SST data.

Spatial context is less critical with high spectral resolution.

Including temporal variation could improve detection in SDO/AIA data.

Abstract

Ellerman bombs (EBs) are observable signatures of photospheric small-scale magnetic reconnection events. The reliable automatic detection of EBs would enable the study of the impact of magnetic reconnection on the Sun's dynamics. We aim to develop a method to automatically detect EBs in H$\alpha$ observations from the Swedish 1-m Solar Telescope (SST) and in SDO/AIA observations using the 1600\r{A}, 1700\r{A}, 171\r{A} and 304\r{A} passbands. We trained models based on neural networks (NNs) to perform automatic detection of EBs. Additionally, we used different types of NNs to study how different properties contribute to the detection of EBs. We find that for SST observations, the NN-based models are proficient at detecting EBs. With sufficiently high spectral resolution, the spatial context is not required to detect EBs. However, as we degrade the spectral and spatial resolution, the spatial information becomes more important. Models that include both dimensions perform best. For SDO/AIA, the models struggle to reliably distinguish between EBs and bright patches of different origin. Permutation feature importance revealed that the H$\alpha$ line wings (around $\pm$1 \r{A} from line center) are the most informative features for EB detection. For the SDO/AIA case, the 1600\r{A} channel is the most relevant one when used in combination with 171\r{A} and 304\r{A}. The combination of the four different SDO/AIA passbands is not informative enough to accurately classify EBs. From our analysis of a few sample SDO/AIA 1600\r{A} and 1700\r{A} light curves, we conclude that inclusion of the temporal variation may be a significant step towards establishing an effective EB detection method that can be applied to the extensive SDO/AIA database of observations. Abstract modified for ArXiv purposes.