Source research

Drones Classification based on Millimeter Wave Radar Cross Section via Deep Learning

Lina Wang, Li Meng, Oluyomi Simpson, Pan Cao · 2024 · 2 citationsRead the paper

The wide use of drones is playing a vital role in the era of low-altitude economy, but meanwhile its misuse could also presents a threat to the privacy, property damage, health and public safety. Therefore, there is a timely need to enhance the capability to detect and recognize the flying drones, which however is challenging owing to the small size, slow speed and low altitude of small drones. In the 5G and beyond, the widely deployed base stations are becoming more and more advanced, especially with the large spectrum like the millimetre wave (mmWave) frequency band. This provides a great potential to turn the network of base stations into a network of mmWave radars for the drones’ detection by leveraging the integrated sensing and communication (ISAC) techniques. This work aims to classify the drones based on their mmWave radar cross section (RCS) real data that will be converted to two-dimensional (2D) RCS images. Thus, 2D Convolutional Neural Network (CNN) is applied to the image sets to achieve the drone classification. The satisfactory testing results verify the proposed drone classification method.

What the full text shows

Shown in the lab

Two-drone classification (M100 and WV4) achieved perfect prediction accuracy. — 100% accuracy (10 runs with random 80/20 train-test split of RCS images from 26GHz to 40GHz)
Three-drone classification (M100, F450, and WV4) achieved high average prediction accuracy. — 82.2% average accuracy (10 runs with random 80/20 train-test split of RCS images from 26GHz to 40GHz)
Seven-drone classification (M100, F450, WV4, P4P, Heli, Hexa, and Y600) achieved reasonable prediction accuracy. — 79.4% average accuracy (Average of three independent simulation tests using RCS images from 26GHz to 40GHz)

Setup: Dataset of 9 drones (8 multi-rotors, 1 helicopter) measured in an anechoic chamber at Aalto University; frequencies from 26 GHz to 40 GHz; angular ranges θ ∈ [-90, 90] and φ ∈ [0, 180] degrees.
How: Raw mmWave RCS data was reshaped into 181x181 2D images representing observation angles. These images were used to train a 2D Convolutional Neural Network (CNN) for multi-class drone identification.
Limitations: Small dataset size (75 total 2D RCS images for the 7-drone study) · Measurements limited to the bottom hemisphere of the drones · Data collected in a controlled anechoic chamber rather than real-world environments

Summarized by Gemma 4 from the open-access full text · 16 Oct 2026. Read it