Incoherent radar spectrum processing using neural network

Abstract

In the high-latitude F-region, the plasma horizontal drift can be comparable to, or greater than the neutral oxygen (O) thermal speed. As a result of the interplay between the ionospheric E×B drift and the O–O+ collisions, the O+ velocity distribution function, fo+(v), deviates significantly from Maxwellian. While the E×B drift velocity increases, the shape of fo+(v) becomes bi-Maxwellian, and then toroidal. The corresponding scattered wave spectrum of the incoherent radar changes from a double peaked, to a triple peaked, and then to a ‘baby-bottle’ shape. The conventional analysis technique, which assumes a Maxwellian, is not suitable under such conditions. Here, we applied artificial neural networks to analyze the radar spectrum corresponding to a non-Maxwellian fo+. A backpropagation neural network was adopted with various structures. The neural network was trained with 3000 sets of spectral density values and tested with another 1000 sets. The performance of the neural network was investigated under different structural configurations and design parameters. It was concluded that: (1) the neural network was tolerant to high levels of noise; (2) its accuracy was acceptable for as few as 10 input nodes; (3) its accuracy was comparable for all line-of-sight directions, and (4) the neural network was not sensitive to the number of hidden layers. The experimental results demonstrate that neural networks can perform well, even with very noisy data, low sampling rates, and wide ranges of ionospheric conditions. Another interesting result contrary to theories propagating through previous literature is that, the accuracy of the analysis is quite independent of the aspect angle (the angle between the radar’s line of sight and the direction of the geomagnetic field). The flexibility, robustness and powerfulness of the neural network approach in processing incoherent radar spectra are further proved by our study.