Background
The Sun often undergoes impulsive disturbances resulting in solar flares, solar energetic particles, and coronal mass ejections. These events and their resulting interactions with Earth’s magnetosphere and upper atmosphere are referred to as space weather. Space weather poses significant hazards to technological systems near Earth: damage to satellites, disruption of communication systems, harm to astronauts, and surges in power grids. To improve space weather forecasts, the National Oceanic and Atmospheric Administration (NOAA) uses automatic segmentation of solar images to label different structures on the Sun every four minutes. The current operational approach, from 2019, segments solar images using a pixel-by-pixel approach without taking advantage of textural clues a human forecaster would— a limitation with detrimental impacts on solar weather monitoring due to mislabeled regions.
Our objective was to use large-scale neural network foundation models to improve the automatic generation of space weather thematic maps, making them more accurate than current state-of-the art applications by factoring uncertainty into their generation.
Approach
A neural network foundation model is any machine learning model which is first trained on a generic task and then refined through fine-tuning with specific narrow tasks. We used the Segment Anything Model (SAM) from Meta AI. We trained an ensemble of ten SAM. Each ensemble was conditioned with a different random state and shown a different subset of our data. Using an ensemble allows us to characterize the uncertainty in the model predictions, increasing their value to human forecasters by indicating which parts of the map can be trusted.
Figure 1: The figure shows the results of training an ensemble for coronal hole detection. “We improved upon the current operational NOAA model, shown in the upper left, using an ensemble that makes predictions like the upper right image from input images shown in the bottom right. More votes indicate the model was more confident.”
Accomplishments
We prototyped a new machine learning algorithm for automated solar image segmentation that qualitatively performs better than the existing operational model. As part of the work, we developed new capabilities for solar image annotation by experts and began preparing a new machine learning ready dataset. This work was presented at the Boulder Solar Day on September 29, 2023. It was also accepted for presentation at the annual meeting of the American Geophysical Union in December 2023.