Abstract
We investigate the use of neural networks (NNs) to automatically identify Ellerman bombs (EBs)—small-scale magnetic reconnection events on the Sun. Using high-resolution observations from the Swedish 1-m Solar Telescope (SST) and degraded data matching SDO/AIA, we evaluate the performance of models incorporating spectral and spatial dimensions. Feature importance analysis reveals that the Hα line wings (approximately ±1 Å from the line center) are the most critical features for identifying EBs in SST data, and that spatial context becomes increasingly important as spectral and spatial resolutions are degraded. For SDO/AIA data, the models showed that the combination of four passbands (1600 Å, 1700 Å, 171 Å, and 304 Å) is insufficient for high-confidence classification, highlighting the need to incorporate temporal variations for lower resolution observations.
Figure 7: Example of EB predictions done by the CNN and FNN SST models over a cutout of the new dataset. The left panel shows the intensity image in the red wing of H α𝛼\alphaitalic_α at +1 Å offset. The middle and right panels show the predictions done by the CNN and FNN models respectively. The colormap indicates the probability assigned by the models to each pixel of being an EB. The white dashed contours mark the 0.1 probability of the CNN prediction and are drawn in all panels to facilitate spatial comparison. An animation of this figure that shows the probability maps over the full time sequence is available in the online material.
