Speaker
Description
Since the development of the LOw Frequency ARray (LOFAR), it has played an important role in the search for variable sources at low radio frequencies. In recent years, numerous types of transient radio pulsations have been identified, including variable x-ray binaries and giant pulses from pulsars. To systematically search for these sources, an automated pipeline TraP was developed. This pipeline evaluates flux densities of detected sources to produce lightcurves for transient candidates.
In transient and variability studies using TraP, the Gaussian fit for each source is often constrained to have the same shape as the restoring beam, effectively assuming all sources are not resolved. However, extended sources spread their emission over a larger area then the beam leading to a systematic underestimation of their flux densities. On timescales relevant in this study, extended sources are expected to be constant in flux. If such sources are selected as variable candidates based on their variability metrics, their variability is most likely due to systematics or propagation effects such as scintillation. A straightforward solution to remove these false positives is to exclude extended sources from the list of candidates. Although separating these populations is not straightforward, filtering out a subset of extended sources can significantly benefit transient searches.
To test this, we aim to train a classification model that identifies and removes sources with a high probability of being extended. Specifically, the model will be designed to filter transient candidates identified by TraP in images produced by the LOFAR Initial Calibration (LINC) pipeline. To remove candidates that have a high probability of being extended, we will test simple but computationally efficient binary classifiers, such as logistic regression, random Forrest or Support Vector Classifier (SVC), implemented using Scikit-learn.
| Talk category | NOVA Network 3 |
|---|
