Robust UAV Classification in Noisy Environments: A Digital Twin-Driven CNN Approach with ROC Analysis

This research presents a systematic evaluation of Deep Learning (DL) models for Unmanned Aerial Vehicle (UAV) classification using Range-Doppler Maps (RDMs) under varying noise conditions. A structured framework is introduced, where synthetic RDMs for six UAV types are generated using Ansys HFSS digital twin simulations. The proposed Convolutional Neural Network (CNN) is trained entirely on noise-free data and evaluated against Additive White Gaussian Noise (AWGN) across signal-to-noise ratio (SNR) levels ranging from –20 dB to 10 dB. The model achieves a validation accuracy of 94.95 % with only 25.7 M parameters, demonstrating strong noise resilience. Receiver Operating Characteristic (ROC) analysis is employed to assess classification performance under noise degradation. This paper establishes a new benchmark for radar-based UAV classification in noisy environments while maintaining state-of-the-art accuracy, providing valuable insights for robust electromagnetic simulation-to-classification pipelines.