Data-driven approach to mid-latitude coherent scatter radar data classification

Abstract

A self-consistent, data-driven approach to classifying data obtained at the ISTP SB RAS mid-latitude coherent scatter radars has been developed. Based on 2021 data, a solution of the problem of automatic data classification is presented without their labeling by an expert and without postulating the number of classes. The algorithm automatically labels the data, determines the optimal number of signal classes observed by the radars, and trains a two-layer classifying neural network of an extremely simple structure. The trajectory calculations use the wave optics method and international reference models of the ionosphere and the geomagnetic field. The model is trained on signals coming from the main lobe of the antenna pattern. During training, to adapt part of the data obtained with improved spectral resolution, it is artificially coarsened to the standard resolution. Each signal class determined by the neural network is interpreted from a physical point of view, using statistical characteristics of the signals belonging to it. The number of classes in the data is demonstrated to range from 23 to 35. The significance of various parameters of the input data is assessed. It is shown that the most important parameters for the classification are the calculated scattering height and the elevation of the trajectory at the scattering point, and the least important are the spectral width of the received signal and the calculated number of reflections from the underlying surface.

Keywords

• decameter radar
• SECIRA
• ionosphere
• automatic classification