Enhanced Drone Classification using Transfer Learning and Optimized RF-Spectrogram

This paper presents important contributions toward state-of-the-art drone classification systems that are deployable in challenging environments. Building on the groundwork laid by existing drone categorization methods, we provide two major contributions: (i) a new processing pipeline that converts RF signals into optimized spectrogram images for computer vision applications, and (ii) an extensive analysis of nine state-of-the-art convolutional neural network (CNN) models, from light to heavy models. Our approach achieves significant improvements in accuracy, efficiency, and real-world deployability over state-of-the-art VGG baseline models. The best-performing model, MobileNetV2, attains comparable or better classification accuracy at very low computational complexity and model sizes. It has 95.0% test accuracy with a mere 2.23 million parameters and 8.52MB model size—highly suitable for deployment in resource-constrained environments. By thorough benchmarking and comparison, we illustrate that chosen CNN architectures, along with an optimized power spectrum image representation, present an effective solution for building deployable and robust drone classification systems in challenging real-world environments.

Index Terms: drone classification, RF signals, transfer learning, CNN architectures, spectrogram, MobileNetV2

Recommended citation: Kaushik A R, Annaamalai U, and Padmavathi S, "Enhanced Drone Classification using Transfer Learning and Optimized RF-Spectrogram," Presented at Conference, July 6, 2025.