Image dataset for the creation of an automatic system for meteor fall detection

Image dataset with sky photos showing the occurrence or non-occurrence of falling meteors. The database comprises 7,000 images in JPEG format -- 3,850 (55%) images show the event of falling meteors, and 3,150 (45%) images show no meteors. Different instruments captured the photos from 2014 to 2023. We used the images to train a deep-learning neural network for an automatic falling meteor detector.

The primary image data sources were the Brazilian Meteor Observation Network (BRAMON -- http://www.bramonmeteor.org), UK Meteor Network (UKMON -- https://ukmeteornetwork.co.uk), and Base des Observateurs Amateurs de Météores (BOAM -- http://boam.fr) repositories.

Folders Structure

We divided the folder structure into two levels. In the first level, we have two folders: RawImages, which holds images with captions stored in the repositories; and CroppedImages, which contains images without the captions (we cropped a band of 24 pixels in the lower part of the image).

In the second level, in each of the previous folders, we have another two folders: meteor, which has images with meteors; and non-meteors, with images without occurrences of meteors.

Naming pattern for the files

The naming pattern in the meteor folder follows the format __.jpg where:

is one of the 3 data sources: bramon, ukmon, or boam.

is the date-time the instrument captured the image in the format yyyymmdd_hhnnss (y:year, m:month, d:day, h:hours, n:minutes, s:seconds).

is an identifier from a specific source to avoid date-time conflicts:

BRAMON: radiant identifier.

UKMON: station identifier.

BOAM: station identifier.

For the non-meteor folder, the naming pattern is __nonmeteor.jpg to avoid homonyms (with the same date-time) and to identify that they are images of non-meteors.