Multimodal Features based Classification of Aerial Non-Cooperative Targets using Machine Learning

Fixed wing aerial targets like drones and loitering munitions which can have long ranges are increasingly becoming a potent threat. There are many limitations in acquiring their target signature particularly as the distance from Radar or other type of sensor increases. It is essential to utilize all possible target features from multiple domains for the purpose of target classification. Modern Radars can provide RCS estimates and track data which can be jointly used to classify aerial targets. In this paper authors have proposed machine learning algorithms to learn targets’ dynamic RCS characteristics along with kinematic characteristics from a sufficiently large dataset and tested their combined efficacy for target classification. By leveraging the strength of different domain features, proposed method system aims to overcome the limitations of individual sensing modalities. Fixed wing aerial targets for classification have been chosen, although the approach can be extended for other types of targets. A unique method for accurately determining aspect angle of targets has been used in this work. This method utilizes automatic dependent surveillance-broadcast (ADS-B) track data. Electromagnetic simulation-based method has been chosen for RCS estimation as the target moves along the flight path. The classification accuracy has been boosted by joint exploitation of RCS and kinematic features along-with derived features. Multiple machine learning algorithms have been tested for classification of targets. The simulation results suggest that XGBoost gives the best classification accuracy. Models’ relative performance have been verified using statistical techniques like Friedman’s ANOVA test.