Abstract Details
| Name: Tejas Kale Affiliation: Cumulus Systems Pvt. Ltd. Conference ID: ASI2016_834 Title : Machine learning techniques for identifying large scale radio sources Authors and Co-Authors : Pratik Dabhade, IUCAA; Madhuri Gaikwad, NCRA Abstract Type : Poster Abstract Category : Instrumentation and Techniques Abstract : Finding giant radio sources is still a process that relies heavily on diligent observation by the human eye. Typical procedure consists of going through large area radio maps to look for diffuse long emission. The results are reliable but slow. Additionally, this process cannot be employed to handle large radio survey data from LOFAR, TGSS, VLASS, etc. Hence, dire need exists to quicken the process by taking advantage of fast computers and advanced mathematical algorithms. But attempts so far to automate it have faced difficulty primarily owing to large, non-uniform sizes and relatively low surface brightness of these giant radio sources. Machine learning (ML) is the science of making computers learn without being explicitly programmed. Learning here refer to the process of fitting the best statistical model of a specified type in which the subject in question (response variable - continuous or ordinal) is modeled on a set of factors influencing the subject (explanatory variables - continuous or ordinal). Herein we will explore the applicability of two popular category of ML algorithms for classification of radio sources. The first is Artificial Neural Networks (ANN) for facial recognition wherein the algorithm makes its best attempt to split a source into mutually exclusive (latent) features. The other category is Logistic Regression, with attributes similar those used in handwriting recognition, in which we define a set of features like grayscale density, object dimensions, etc. whose distributions are useful for segregating different types of sources. In both the approaches, using a training dataset, the algorithm learns to associate a joint distribution of these features with each type of source. A test data is then used to determine the rate of misclassification of the model for each type of source to comment on its overall efficacy for wider application. For this purpose we choose NVSS radio maps to identify large scale radio sources. NRAO VLA SKY SURVEY( NVSS) is one of the best all sky radio survey till date. NVSS is a 1.4 GHz continuum survey covering the entire sky north of −40 deg declination (Condon et al 1998). Its images are made with a relatively large restoring beam (45 arcseconds FWHM) to yield the high surface-brightness sensitivity needed for completeness and photometric accuracy which is very essential in context of our project. Their RMS brightness fluctuations are about 0.45 mJy/beam = 0.14 K (Stokes I) and 0.29 mJy/beam =0.09 K (Stokes Q and U). The RMS uncertainties in right ascension and declination vary from < 1 arcseconds for relatively strong (S > 15 mJy) point sources to 7 arcseconds for the faintest (S = 2.3 mJy) detectable sources with a completeness limit is about 2.5 mJy. |