IN51A-1781
Machine Learning in Ionospheric Phenomena Detection Using Passive Radar

Friday, 18 December 2015
Poster Hall (Moscone South)
Soubhik Barari1, Victor Pankratius2 and Frank David Lind2, (1)Tufts University, Medford, MA, United States, (2)MIT Haystack Observatory, Westford, MA, United States
Abstract:
This work describes an approach to automate ionospheric feature detection in passive radar data using a tunable pipeline of Python-implemented algorithms for detection and classification. In particular, our detector is tuned to capture E-region irregularities and various other events such as meteors, aircraft, and ambiguities that result from poor transmission of signals or noise interference. The detection stage applies to passive radar images with pixels normalized to a defined value range. To separate the background, we apply a thresholding value and an area cuttoff to keep regions with connected pixels of a minimum size; for each particular image, these parameters can be determined algorithmically in two ways through our ExplainedEntropy (EE) and MaximumRegionArea (MRA) techniques. EE identifies the smallest set of regions that explain the most entropy of the image. MRA sets the area threshold to be a function of the largest region size. The classification stage picks up on these detected areas and applies neural networks and random forests to the image feature space. This way we are able categorize images based on their scientific content and make them searchable for scientists.

A training set of real radar images was available to evaluate our approach and its adaptivity. Based on these labeled real images, we also evaluated the robustness of the detection with enhanced set of perturbed images that were generated through a model-based simulator. The simulator also allowed for controlled experiments in the amount of perturbation and noise added, to precisely characterize the operation ranges of our machine learning algorithms. We will discuss the performance of the algorithms and potential scientific applications.

Acknowledgements. We would like to acknowledge support from the NSF ACI-1442997 (PI V. Pankratius).