Segmentation of coronal holes with a convolutional neural network

This repository contains the source code to reproduce the coronal holes segmentation model from the paper Segmentation of coronal holes in solar disk images with a convolutional neural network (published in MNRAS).

Demo

Try a demo running directly in the browser to see how the model will process solar disk images you will feed to it.

Installation and quickstart

Clone the repo

git clone https://github.com/observethesun/coronal_holes.git

or download it as an archive.

For quickstart you need to have an archive of SDO/AIA 193 A jpeg images in 1K resolution that correspond to annotated coronal holes in date and time of observation and check that all required libraries are installed.

Content

The repository is organized as follows:

• src folder contains the neural network architecture, methods for image loading and processing and classes that handle large datasets.
• notebooks folder contains tutorials that explain how everything works. More detailed, you may need
  • train_segmentation_model.ipynb to reproduce the model training procedure
  • run_segmentation_model.ipynb to test the model on SDO/AIA images
  • convert_images.ipynb to convert .jpeg images to more memory efficient .blosc file format.
• data folder provides a daily archive of coronal holes (CHs) in .abp file format. CHs were obtained from SDO/AIA 193 A images in 1K resolution for the period 2010 to 2018. These .abp files are targets for the segmentation model. Visit http://observethesun.com/ to look at the distribution of CHs over the Sun.
• requirements.txt contains a list of python libraries required to run the code.

What is an .abp file format

.abp is a text format used at the Kislovodsk Mountain Astronomical Station to describe active regions (in particular, coronal holes) isolated in solar disk images.

All .abp files have similar structure:

• Filename contains date and time of the solar disk observation.
• First line of a file contains 7 numbers. First three numbers are x- and y-coordinates of the solar disk center and its radius given in pixel units. Last numbers are not important for CHs.
• Second line is not important for CHs.
• Third line contains a number of active regions.
• The rest lines are given by pairs of lines one for each active region. Pairs are organized as follows:
  • first line contains 7 numbers, where the first one is an index of active region within the current file, second one in a number of pixels occupied by active region. Following numbers are not important for CHs.
  • second line contains triples x, y, c one for each pixel within the active region. Here x, y are coordinates of pixel belonging to active region and c is a flag whether this pixel is inner (c=2) or edge (c=1) for the active region. Coordinates are in pixel units.

Where to get SDO/AIA images

We suggest http://jsoc.stanford.edu/ for massive queries. For single images with preview consider http://suntoday.lmsal.com/suntoday/.

Why .blosc not .jpeg

During the neural network training we read images from disk multiple times. However, I/O operation are known to be slow. To speed up the process we suggest to keep data in optimized formats that can be operated faster. Using .blosc format one can have a speed benefit up to several times. Read more at http://blosc.org/.

Citing this work

Illarionov E., Tlatov А., 2018, MNRAS, 481, 4.

Using provided .abp files in your research please cite

Tlatov A., Tavastsherna K., Vasil’eva V., 2014, Solar Physics, 289, 4.