Automated Solar Activity Prediction (ASAP)

What is ASAP?

Automated Solar Activity Prediction (a.k.a ASAP) is a system proposed to predict flares (figure below shows the currently completed part) and CMEs in near real-time. ASAP currently works for flare prediction, and further work on CMEs continues.

ASAP flare-prediction pipeline: SDO/HMI magnetogram and continuum images feed sunspot and active-region detection, which are grouped and classified, then combined with GOES X-ray flux data in a machine learning system to produce flare predictions. SDO/HMIMagnetogramImages SDO/HMIContinuumImages Detect Activeregions Detect Sunspots Group and ClassifySunspots MachineLearningSystem FLAREPrediction GOES
ASAP's flare-prediction pipeline: SDO/HMI magnetogram and continuum images feed sunspot & active-region detection, which are grouped and classified, then combined with GOES X-ray flux data in a machine-learning system to produce flare predictions.

How do we predict flares?

We use SOHO/MDI (SDO/HMI) Continuum and Magnetogram images to detect sunspots and find their McIntosh classifications. These classifications are fed to the learning models, which were previously created in [Qahwaji and Colak, Solar Physics, 2007], to provide real-time prediction for the possible occurrence of flares. The flare prediction system can be described in three stages:

1

Sunspot Grouping

  • Detection of sunspot candidates from MDI continuum images using morphological image processing algorithms.
  • Detection of active region candidates from MDI magnetogram images using morphological image processing algorithms. The MDI magnetogram images show the magnetic fields of the solar photosphere, with black and white areas indicating opposite magnetic polarities. These areas are detected separately and combined afterwards to determine the active region candidates.
  • Applying region growing to combine sunspot and active region candidates.
  • Using neural networks to combine regions of opposite magnetic polarities in order to determine the exact boundaries of sunspot groups.
Sunspot grouping pipeline from continuum and magnetogram images to detected sunspot groups
Continuum and magnetogram images are processed to detect sunspot and active-region candidates, then combined and grouped by a neural network into the final sunspot groups (shown false-coloured).
2

McIntosh-Based Classification

  • Extracting local features from every sunspot in every group using image processing and neural networks: length, tallness and area; and using neural networks to decide the type of penumbra (Mature or Rudimentary) and whether the sunspot is Symmetric or Asymmetric.
  • Extracting features from each sunspot group using image processing — length, largest spot, polarity and distribution.
  • Applying all the extracted features to a decision tree to determine their McIntosh classification.
Sample ASAP output showing McIntosh-classified active regions and flare probabilities
Sample ASAP output: each active region is labelled with its McIntosh classification (e.g. EAO, CAO) alongside M- and X-class flare probabilities.
3

Flare Prediction using Neural Networks

  • The publicly available sunspot group catalogue and the solar flare catalogues were investigated to associate flares with the sunspots that caused them. The association is determined based on the location (i.e., same NOAA number) and timing information.
  • A neural network is optimised and trained using this association information. The input for the networks is composed of sunspot group classification (McIntosh) and sunspot areas — one network for determining flaring possibility, one for determining probability of flare type.
  • The two neural networks are combined to produce a hybrid system that gives the flaring probability of each sunspot group and its flare-intensity probabilities.

How Accurate Are Our Predictions?

Accuracy depends mostly on the accuracy of each stage above. In total, this means a success rate of ~70% on final flare prediction when we combine all stages.

95% Sunspot grouping accuracy
80% Sunspot group classification accuracy
90% Flare prediction accuracy, given correct classification
70% Overall end-to-end success rate