Naslov
"Automated Classification of LSST Images Using Convolutional Neural Networks"

Autori
Mrakovčić ; Karlo

Vrsta, podvrsta i kategorija rada
Ocjenski radovi, diplomski rad, diplomski

Fakultet
Fakultet za fiziku

Mjesto
Rijeka

Datum
06.09

Godina
2022

Stranica
92

Mentor
Dominis Prester, Dijana

Neposredni voditelj
Željko Ivezić

Ključne riječi
strojno učenje ; fotometrija
(machine learning ; photometry)

Sažetak
With its 3.2 gigapixel camera, the Vera Rubin observatory can produce im- ages of up to 10 million changing celestial sources per night. Its first principal project, the Legacy Survey of Space and Time (LSST), will be a multi-band large- area time-domain astronomical survey. Because the data is impossible to classify and analyze in the conventional way, the telescope will need to take a different approach in order to generate quality scientific products. Because of the volume of data collected (20 TB each night), accurate data categorization and false positive detection automation must be created. Images of three types of sources were generated to approximate anticipated LSST images: ”stars, ” ”trails, ” and ”dipoles.” Unsupervised learning was utilized to automate the categorization of generated LSST images using a combination of convolutional autoencoder and Kmeans. Hyperparameter search was done on HPC Bura to explore hyperparameter space, and cMetric, a measure for evaluating different models, was devised. The accuracy of 91.2 percent was reached, presenting us with a promising tool that can be integrated one day into LSST pipeline.

Izvorni jezik
Engleski

Znanstvena područja
Fizika

POVEZANOST RADA