Ground-based gamma-ray astronomy studies very energetic radiation of galactic and extragalactic origin by specially designed telescopes, the so-called Imaging Air Cherenkov Telescopes (IACTs). With this technique, gamma-rays are observed on the ground optically via the Cherenkov light emitted by extensive showers of secondary particles in the air when a very-high-energy gammaray strikes the atmosphere. Gamma-rays of such energies contribute only a fraction below one per million to the flux of cosmic rays, most of which are protons. Nevertheless, being particles without electric charge they can be extrapolated back to their origin, which makes them the best “messengers” of exotic and extreme processes and physical conditions in the Universe. That is why particle identification (gamma-ray discrimination against the cosmic-ray background) is an essential part of data analysis for the IACT technique.
Online gamma-hadron separation using CNN
Here you can find a prototype of the Astroparticle CNN client developed for gamma-astronomy tasks. This prototype provides access to on-line analysis of the gamma/hadron separation using convolutional neural networks developed as part of the GRADLCI project. The Monte Carlo events of the TAIGA-IACT telescope are used to operate this prototype. The developed convolutional neural networks get the probability of the gamma reconstruction for each event as a result. You can also check yourself if you can make gamma/hadron separation using the telescope image (instruction is below).
Instruction how to define gamma-event using telescope image
A shower image is fitted as an ellipse. The ellipse is characterized by its axes and has parameters length, width, distance and the angular miss-alignment of the major axis α. In comparison with hadron showers gamma-ray ones have more elliptic shape, less width and major axis pointed to the source.
If the telescope is pointed at a known source, the source is positioned at the center of the telescope camera field of view. Therefore, the axes of gamma-induced images are pointed to the center of the camera field of view.
However, real data are taken mostly in separate 20-minute runs using the so-called wobble mode. It’s an observation mode in which the source is not in the center of the camera.
The wobble mode allows simultaneous source observation and hadron background estimation by wobbling the telescope tracking position around the source location. This eliminates the need for special off-source observations and consequently doubles the amount of available on-source time.
In this mode for the TAIGA telescope, the source direction is positioned 1.2 degree in declination relative to the center of the camera field of view. The sign of the offset is altered in successive runs to reduce systematic effects.
To demonstrate the operation of the neural networks there five datasets have been prepaired, each contains 10 telescope events with the event information: event ID, particle energy (TeV), particle type, image. You can also download these datasets for your further research:
Process you own dataset
To test the neural nets on your own events just upload your dataset.
The file structure and an example how to print the structure of the HDF5 file using python are described below you can find here.