Speaker
Description
Introduction: In this study, we applied Explainable Artificial Intelligence (XAI) techniques (LIME and SHAP) to a Convolutional Neural Network (CNN) trained to classify Helioseismic and Magnetic Imager (HMI) magnetograms according to the presence/absence of an Hsx sunspot group. Our focus is on Solar Satellite Images (SSIs), where the data constitute a multimodal mixture of regular image collections and associated semi-structured Solar Flare textual information that provides critical context and labels.
Methodology: The dataset employed in this study was obtained from the Multiple Data-source Solar Extraction, Transformation, and Load (MS-ETL) pipeline, which aggregates raw HMI image collected by the Solar Dynamics Observatory (SDO), incorporating metadata such as timestamp, solar coordinates, active-region identifiers, and associated NOAA classifications. Each image corresponds to a line-of-sight HMI magnetogram with a mapped resolution of 224x224 pixels. MS-ETL performs several preprocessing steps (disk detection, limb darkening correction, coordinate referencing, and cropping) to produce standardized solar patches centered on active regions. These cleaned and uniformly formatted images serve as an ideal basis for DL workflows. In this work, we extracted all samples labeled under the presence or absence of the Hsx sunspot classification (canonical category of the McIntosh - Mount Wilson). The dataset is therefore divided into two classes: (a) no-Hsx: magnetograms without an Hsx-type sunspot group, and (b) with-Hsx: magnetograms containing an Hsx-class sunspot, characterized by a compact, well-formed magnetic core. To interpret the CNN’s predictions and assess whether it identifies physically meaningful solar features, two complementary XAI techniques were applied: LIME and SHAP. These methods provide local explanations for individual predictions, but differ significantly in methodology and interpretive power.
Results: We found that LIME identifies influential regions in close proximity to the localized physical structures of the sunspots, effectively isolating the umbra and penumbrae as the primary drivers of the classification. In contrast, SHAP attributions display a more global distribution, indicating that the neural network's decision-making process is also influenced by background distortions and the geometric boundaries of the solar limb.
Although both methods confirm the model's sensitivity to magnetic complexity, the disparity in their output suggests that LIME is more effective at capturing local morphological features, whereas SHAP uncovers the model's reliance on broader contextual and structural artifacts within the satellite imagery. These findings demonstrate that utilizing a multi-method XAI approach is relevant for identifying unintended model biases and for ensuring the transparency to DL methods.