Evaluating Solar Imaging Feature Extraction Techniques for Enhancing Space Weather Prediction with Deep Learning Models

Many machine learning models have provided significant results in predicting the geomagnetic activity quantified by Earth-measured geomagnetic indices. For instance, one such model is the SERENADE model that provides probabilistic forecasts of the Kp index, days ahead solely from solar imaging. It consists of three modules combining convolutional, recurrent, and linear neural network layers that first extract the important information contained in the input solar imagery and transform them into an intelligible forecast. To improve the performance of this model, we evaluate solar-imaging-adapted dimensionality reduction techniques that extract the features from the images and can therefore be used as the first layer of the forecast model. We use a solar imagery dataset formatted specifically for machine-learning research (SDOML). We applied the Principal Component Analysis method and trained AutoEncoders and Variational AutoEncoders (VAE) targeting several reduced dimensions. We consider the convolutional GoogLeNet method, which was pre-trained on the ImageNet dataset, as a baseline for our comparison. We analyze the information retained by the extracted features in terms of solar activity physical parameters and find high correlations between the latter and the the reduced representations of the images, with the VAE results standing out. In addition, we re-train the SERENADE model to predict the daily maximum of the Kp index two days in advance using the extracted features by the new dimensionality reduction methods as input to the model. We first use the same hyperparameters that were optimized for the GoogLeNet model and obtain more stable predictions using the dedicated solar imaging feature extractors than when using the baseline model, specifically in the VAE case. Furthermore, when fine-tuning SERENADE's hyperparameters to the VAE model, the predictive performance of the model was enhanced, notably during geomagnetic storms, which indicates that the use of adapted feature extractors could improve the geomagnetic activity forecasting.