Comparison of Solar Imaging Feature Extraction Methods in the Context of Space Weather Prediction With Deep Learning‐Based Models

Recently, many machine learning‐based models have been developed to predict geomagnetic activity several days in advance directly from space‐borne solar imaging. To better understand and improve these models, we compare dimensionality reduction techniques to extract abstract features from solar images for space weather‐related downstream tasks. We applied the principal component analysis method and trained autoencoders and variational autoencoders (VAE) targeting two reduced dimensions: 1024 and 256, from the original 512 × 512 images. These methods are compared to a baseline convolutional GoogLeNet model pretrained on the ImageNet data set. We first ensure that the models have been properly trained by evaluating their compression and reconstruction capabilities. We then analyze the latent spaces and find high correlations between the reduced representations of the images and solar activity physical parameters, with the VAE showing the highest potential. In addition, we retrain the SERENADE deep learning model to predict the daily maximum of the Kp index 2 days in advance from the features extracted using each of the dimensionality reduction methods. We find that the predictive accuracy of the new models is similar to that of the baseline model, without any new search for optimized hyperparameters. However, the predictions obtained with the feature extractors dedicated to solar images are much more stable than those obtained with the model pretrained on ImageNet. Furthermore, the analysis indicates that with additional fine‐tuning of SERENADE's hyperparameters, the adapted feature extractors could improve predictions during geomagnetic storms, thereby facilitating future advances in days‐ahead geomagnetic activity forecasting.