Skip to content

Latest commit

 

History

History
69 lines (46 loc) · 3.15 KB

README.md

File metadata and controls

69 lines (46 loc) · 3.15 KB

SEgmentation of polAR Coronal Holes (SEARCH)

Using unsupervised learning to SEARCH for polar coronal holes in synoptic EUV images.

Contents

  1. Data Reduction Pipeline
  2. Unsupervised Learning for CHs and ARs Segmentation

Data Prep

Plan: from now until CoolStars 20.5 (March 2-4)

Objective: Generate multi-wavelength (171A, 195A, and 304A) synchronic maps using SoHO/EIT and STEREO/EUVI data for the period 2010 (included) to 2015 (excluded).

  • Step 1: Wavelet-enhancement of the imagery for improvised contrast (dark coronal holes, bright active regions):

    • Translation of pre-existing IDL code to Python. This would allow us to be consistent with existing wavelet-enhanced satellite data.
    • Compare results to original paper on enhancement method.
    • Compare EUVI enhanced imagery to existing database (at APL?)
    • Alternative methods have been explored by Ajay. Perform comparisons.

  • Step 2: Combination of maps from three different vantage points (homogenized and wavelet-enhanced):

    • Address issues with SunPy for this step (if some still remain). We previously experienced missing patches where there should have been data.

  • Step 3: Generate synchronic maps:

    • Generate synchronic map at 195A and compare to PSI database synchronic maps at 193A/195A. Maps should be consistent.
    • Once imagery is satisfactory, generate synchronic maps at 171A and 304A. Assess quality (though no direct comparisons are possible).

Segmentation

Plan: from now until CoolStars 20.5 (March 2-4)

Objective: Identify CHs and ARs boundaries in synchronic maps through unsupervised machine learning methods (W-net, K-means).

  • Step 1: W-net: Train and test with higher resolution single wavelength (193A/195A) imagery, ideally without downsampling synchronic maps.

  • Step 2: W-net: Optimization for CHs and ARs segmentation. As of now, the W-net has been used out-of-the-box, i.e., without modifications.

    • Modify depth of the U-nets.
    • Other tweaks (TBD).

  • Step 3: Once multi-wavelength data is ready, repeat training and testing of K-means and W-net using:

    • Single-wavelength images (i.e., single channel as input)
    • Multi-wavelength images (i.e., multiple channels as input)
    • Study how the boundaries change depending upon the inputs.

  • Step 4: Study properties of polar CHs:

    • Area
    • Polarity

  • Step 5: Matthew has suggested another approach (hierarchical clustering?) To be explored.

Plan: beyond CoolStars 20.5 (March 2-4); Note: No specific timeline yet.

Objective: Transition to synoptic maps. Two different datasets are to be considered: SoHO/EIS (22+ years) and SDO/AIA (11 years, more recent data, higher resolution).

  • Step 1: Generalization:

    • Use synoptic maps as inputs into algorithms trained on synchronic maps.

  • Step 2: Training and testing of W-net and K-means using single- or multi-wavelength synoptic maps:

    • First set: SoHO/EIT data
    • Second set: SDO/AIA data
    • Compare predictions to algorithms trained on synchronic maps.

  • Step 3: Study properties of polar CHs as the solar cycle evolves:

    • Area
    • Polarity