Chapter 10 - Multiobjective optimization design procedures for data-driven unmanned aerial vehicles automatic target recognition systems

Abstract

Within unmanned aerial vehicle (UAV)-related ongoing research topics, automatic target recognition (ATR) is acquiring increasing relevance. This is due to the easiness, in terms of price and accessibility, with which it is possible to acquire UAVs. ATR systems are responsible for detecting aerial vehicles automatically using radar signals. To this end, multiobjective optimization design (MOOD) procedures can be employed for improving the performance in the presented task. This chapter presents an MOOD procedure for ATR system development composed of three steps: (1) the development of a multiobjective problem, where radar data acquisition of three different UAVs is used to train a classifier in terms of accuracy, precision, and sensitivity; (2) the application of an evolutionary multiobjective optimization algorithm, which retrieves a set of nondominated solutions; and (3) the employment of multicriterion decision making techniques for selecting a final preferred solution. For experimental purposes, two scenarios are evaluated on a new real-world UAV radar signal data set, both discriminating different types of UAVs. The first analysis presents the comparison of two different classification techniques, the multilayer perceptron (MLP) and the support vector machine (SVM). The second analysis presents the comparison of three different multiobjective optimization algorithms, nondominated sorting genetic algorithm version II (NSGA-II), multiobjective differential evolution with spherical pruning (spMODE- II), and multiobjective evolutionary algorithm based on decomposition (MOEA/D), used for finding the best possible trade-off for the automatic target recognition system. As a conclusion, the UAV radar signal dataset is made available for future studies, and current results indicate the superiority of SVM over MLP, and of both spMODE-II and MOEA/D over NSGA-II.