Segment-based feature proposal for the morphological classification of T Tauri star light curves

Abstract

T Tauri stars are young stellar objects that exhibit a wide range of morphological variability in their light curves, product of multiple physical processes. Numerical features can be designed to identify the characteristics of these brightness variations. The identification process is automatized in the literature through the use of machine-learning algorithms. This Master thesis aims to utilize supervised machine-learning algorithms for morphological classification of T Tauri stars with a set of proposed fea- tures based on segmentation strategies. The features are tested on an ex- ternal data set for evaluation and the best parameters for classification are discussed. We evaluate the features through the Welch t-test, the Mann-Whitney U-test and the Levene test for equal variance. After the testing, seven algo- rithms are trained with light curves from the Orion star formation complex obtained from the TESS project. The algorithms are subject to sequential reduction of feature space, hyper parameter grid search and recurrent im- portance calculations to optimize classification results in terms of F1 score and Cohen kappa. The optimized algorithms are then applied to a sample of over 2000 hand-classified confirmed T Tauri stars. In this work, we propose 61 features based upon robust statistics, pseudo- time-series analysis and auto-correlation measurements. These features were utilized in 13 implementations of binary and multi-class classifiers and opti- mized taking advantage of a high-performance cluster. We implement statis- tical tests as feature evaluation and light curve filtering strategies innovative to astronomical feature design. The highest achieving features on the test- ing data set were analyzed individually and were conceptually connected to physical processes, signal crowding and systematic effects. The algorithms obtained F1 scores higher than 0.4 for all classes with maximum feature dimension of 10. This work contributes a new set of useful features that consistently achieve high importance when compared to features used in the litera- ture. Innovative feature design, evaluation stages and algorithm optimiza- tion pipelines were implemented.