SOCIS 2019 Ideas Information
Main description is available at the OA page for this project. Are you looking for the GSoC 2020 version of this idea?
There's no requirement on solar physics knowledge. You will learn a lot about solar active regions and flares on the way. You will need to be familiar with python, image processing and machine learning.
Sunspotter collected lot of data on two campaigns:
Each dataset contains information about the date and location of each active region used in the project and how it was classified by the volunteers. The volunteers were simply asked which image from a set of two seemed more complex, and by an Elo rating system provided a score.
The All-clear dataset was also used on a comparison on flare forecasting.
This project will use different aspects of SunPy:
- Fido to find and download data
- Query solar features archives like HEK
- Work with coordinates
- Visualise data
while using them we may find new needs that the above is not covering and hopefully we can improve them on the way. For example, the idea about designing a SunPy search events mechanism based on the images using HEK or HELIO, could provide information from the available catalogues about the images we are interested (how many spots had that active region? when did it flare? how long did it live? ... ). This could be added as to the region of interest class.
Both dataset have been obtained using SMART (check the other SOCIS project about it) as a segmentation algorithm. Then each image annotated with information of the flaring activity a day or two days later.
You will need to have some background on image processing to understand whether the segmentation may have affected the image shown to the volunteers. For example, on the 14 years dataset, all images are scaled to the same size.
This is where you can bring most of your knowledge! We want to train a model using both these datasets and test it with data from SDO (There will be some normalisation to do as SDO has better resolution). But you should be who suggests which methods to try and when.
Essentially, it's believed that an active region has more chances to flare if it looks more complex. Does this data show that?
- Extract the dates, coordinates and file names from the database and obtain the segments directly from the fits files.
- Graph classification scheme vs flares (data available on the database)
- Unsupervised clustering on the images and or AR properties as detected by SMART
- Neural Network model to produce a classification based on the database ranking.
Main description is available at the OA page for this project.
