Researchers use deep learning and acoustic imaging to detect drones from sound

Scientists have developed a machine learning approach that turns microphone arrays into precision locators for unmanned aerial vehicles by analyzing the acoustic signatures drones produce. Rather than estimating the direction to a sound source as a single angle, the method treats audio detection as a spatial mapping problem, training neural networks to identify and segment regions where active sound originates in three-dimensional space.

How the acoustic imaging works#

The system begins with a custom 24-microphone array that performs delay-and-sum beamforming — a signal processing technique that concentrates on sounds coming from different spatial directions. This generates "beamformed energy maps" that represent both azimuth (horizontal angle) and elevation angles in a 360-degree field.

A modified U-Net architecture, trained on frequency-domain versions of these acoustic maps, then learns to recognize regions where drone noise is present. The network essentially performs semantic segmentation on the audio landscape, marking which spatial directions contain active sound. After the neural network identifies these activated regions, the system computes centroids to produce precise direction-of-arrival estimates.

Dataset and validation#

The researchers trained their model on real-world recordings of a DJI Air 3 drone, collected across multiple dates and locations with synchronized video and GPS flight logs. This ground-truth alignment allowed them to generate binary supervision masks for training.

The team also tested their approach on the DCASE 2019 TAU Spatial Sound Events benchmark, demonstrating that the acoustic segmentation framework generalizes beyond drone detection to broader sound event localization and detection (SELD) tasks involving multiple classes of sound sources.

Technical advantages#

A key strength of the method is its array independence: because the network operates on beamformed energy maps rather than raw microphone signals, it can adapt to different microphone configurations with minimal retraining. The approach also addresses class imbalance in training data using the Tversky loss function, which improves performance when acoustic events are sparse or unevenly distributed.

By reformulating drone localization as a segmentation problem rather than discrete angle regression, the model achieves improved angular precision compared to traditional sound source localization methods. The research suggests a new paradigm for dense spatial audio understanding with potential applications in airspace monitoring and security.