Deep Learning Models for Signal Classification using Allen Telescope Data.
The Search for Extraterrestrial Intelligence (SETI) institute was founded to collect and analyze radio signals from space in an attempt to find clues of advanced civilizations that may perhaps exist elsewhere in the cosmos.The prominent goal of this paper is to appropriately categorize the Allen Telescope data SETI radio waves which were translated into two-dimensional spectrogram representation. This SETI spectrogram classification task is so challenging due to the fact that majority of the classes are non-separable, we propose the use of pre-trained Convolutional Neural Network (CNN) models to classify these spectrograms. The contribution of this paper to the field is the exploration of various CNN models, ensemble models, and hybrid models combining CNNs with machine learning classifiers to improve the accuracy of classification. Amongst different deep learning model variants, the Inception-ResnetV2 CNN hybrid model with ADAM optimizer was found to be efficient in classifying Allen Telescope data with the highest accuracy of 91.44%. This is a significant finding and highlights the potential of using hybrid CNN model which harnesses the best traits from two CNN models in efficiently classifying Allen Telescope SETI radio signals.
--> The data is available on the official SETI website. The four files have seperate train and test files.
--> All the models have been provided in the "Models" folder. It also contains the preprocessing and encoding part.