Source Artefact Classification in Interferometric Images using - - PowerPoint PPT Presentation

Source Artefact Classification in Interferometric Images using Machine Learning

Arun Aniyan

SKA SA & Rhodes University Cape Town , South Africa

Motivation

The field around 3C147, at 5 million dynamic range using 21cm JVLA data.

Smirnov et. al 2015

Motivation

• 1. Improving reliability of source

finders and thus creating better source catalogues. 


• 2. Reducing the steps in the reduction

Objective

Distinguishing real sources from artefacts in radio interferometric images in order to make reliable sources catalogs

Dataset Generation

Simulation pipeline with Meqtrees Noordam & Smirnov , 2010 http://meqtrees.net/

Dataset Generation

Generate sources in random positions Induce DDE + pointing errors

Generate Images for “n” hours of observations

Extract sources with PyBDSM

Cross match with known sky model & PyBDSM model

Extract PyBDSM + generated features

Dataset Generation

Generated simulated skies for the JVLA in C-configuration using L-band Generated Images for observational periods from 1 hr to 25 hrs Ran PyBDSM to get catalogues of sources and artefacts with their features.

Feature Extraction

• 1. Flux features


Total flux, Peak flux

• 2. Axis-Angle features


FWHM of deconvolved major axis etc

• 3. Nearest bright source features


Distance to nearest bright source

Total 28 features

Results

Classifier Accuracy Recall Decision Tree

88.67 92.05

KNN

95.92 99.94

Random Forest

95.22 99.18

Naive Bayes

82.04 84.75

Feature Analysis

Feature Analysis

Features rich in discriminatory power

• 1. Nearest bright source features

2 . Flux features
 


Conclusion and Future Direction

High accuracy Classification of Sources and Artefacts Identification of useful features for classification Development of Convolution Neural Network

arun@ska.ac.za

Thank You !