Improved Solar Coronal Mass Ejection Prediction with Multimodal Machine Learning
Given Earth’s increased dependence on technology – for communications and other critical tasks, and threats to such technologies posed by natural phenomena, we must innovate in order to protect our technology infrastructure now and for future generations. One of the most significant natural threats to our technology infrastructure is the occurrence of coronal mass ejections (CMEs). These high-energy outbursts from the Sun can have severe impacts on Earth’s communications and electrical systems, potentially leading to widespread disruption and damages culminating in the trillions of US dollars.
To address this concern, we build on existing literature to propose a novel approach that leverages the power of multimodal machine learning, integrating physical features and Hu Moment image processing. With a Support Vector Machine, stratified cross validation, and grid search, the model has achieved a True Skill Score of 0.72 ± 0.17, 48 hours ahead of peak event time. This appears to be a roughly 30% improvement when compared with existing literature, which trains with the same data sample set, physical features, and processing techniques.
This research suggests the promising potential of melding image processing and physical measurements for prognosticating CME incidents. That is, the use of spatial statistics, processed from Hu Moments, can explain variance in an outcome beyond those available through traditional aggregate point-source statistics. The inherent advantage lies in the ability of spatial statistics to illustrate complex interactions and relationships in data that have spatial dependency, effectively capturing heterogeneity across the surface of the sun.
For more information, please see our website, which visualizes the predictions in real-time.