Radar-Based UAV Classification: A Micro-Doppler and Deep Learning Approach

Micro-Doppler signatures provide a unique way to characterize the motion of flying objects based on their radar reflections. In this study, we leverage a dataset developed by DIAT to classify six distinct classes of flying objects—three long-blade rotors, three short-blade rotors, birds, bird+mini-helicopter hybrids, drones, and RC planes—using deep learning techniques. We propose and compare two convolutional neural network (CNN) architectures: a custom Deeper CNN and a transfer learning-based EfficientNetB3 model. The Deeper CNN achieved a test accuracy of 91.30%, while the fine-tuned EfficientNetB3 model outperformed it with a test accuracy of 97.10% and a validation accuracy of 97.74%. Our results demonstrate the efficacy of transfer learning for micro-Doppler classification and highlight its potential for real-world radar-based object detection applications.