Abstract
Understanding the evolution of the universe, such as its expansion and the formation of large structures, is a key issue in modern cosmology. To do this, it is essential to recognize the various celestial bodies that allow us to probe and model the distant universe. Type Ia supernovae are particularly interesting because the explosion mechanism is similar for all of these stars. This characteristic allows us, among other things, to deduce the luminous distance of the exploded supernova. Moreover, since these explosions are exceptionally bright, they can be seen at great distances. However, detecting this type of star is not an easy task, as many other celestial bodies exhibit similar behaviors. It is therefore necessary to implement robust machine learning methods to accomplish this task.The analysis of these objects is done using only photometric information, that is, the amplitude response of a celestial object in a few spectral bands (e.g., blue, visible, near-infrared). Observations are carried out over a given period, and we can then construct a time series, called a "light curve," which represents the amplitude values of an object in a specific band over time. Currently, these light curves have been the subject of numerous studies, particularly due to their dissemination during challenges like PlAsTICC 2018. However, the classification results remain insufficient for carrying out astrophysical science.During the thesis, we propose to extend this work by using images of celestial bodies at different times directly, that is, sequences of images centered on the object of interest. The goal is to construct a deep learning-based model capable of extracting contextual information and thus improving performance compared to results obtained from the use of light curves. Implementing such a model is a complex task because astrophysical data is far removed from the optical images used in machine learning and signal processing literature. These images are intrinsically very noisy and have a high dynamic range. Additionally, the database is subject to the problem of mismatch. Indeed, the statistics of the training and test databases are different. This phenomenon is explained by the fact that the observations making up the training database are primarily "nearby" astrophysical objects, which implies that their spectrum is less redshifted than distant objects. This redshift phenomenon drastically alters the perceived image and, therefore, the shape and nature of the images in the test database.As a result of this thesis, we first proposed a novel approach based on deep learning for classifying different types of space objects directly using images. We named our approach ConvEntion, which stands for CONVolutional attENTION. It is based on convolutions and transformers, which are new approaches for the treatment of astronomical image time series. Our solution integrates spatiotemporal features and can be applied to various types of image datasets with any number of bands. In this work, we solved various problems the datasets tend to suffer from, and we presented new results for classifications using astronomical image time series with an increase in accuracy of 13% compared to state-of-the-art approaches that use image time series and a 12% increase compared to approaches that use light curves.As a second part of the thesis, we proposed an end-to-end semi-supervised image time series representation learning framework that leverages self-supervised approaches to reduce the effect of class imbalance with data augmentation, minimize intra-class variance, and improve results on small datasets without the need for extra labeled data. The results of the proposed Semiconformer model were compared to a conventional model and were found to achieve an accuracy of 82.18% and an F1 Score of 75.33%.