ITU ARIS Lab. Acoustic Drone Dataset

CITE -> İ. M. Muhacıroğlu and T. Akgül, "Deep Learning Based Acoustic Drone Detection," 2026 34th Signal Processing and Communications Applications Conference (SIU), İstanbul, Turkiye, 2026, pp. 1-4, doi: 10.1109/SIU71813.2026.11636719. keywords: {Modeling;Acoustic;Audio Processing;CNN;CRNN;Deep Learning;Drone Detection;LSTM;Mel Spectrogram}, <- CITE A two-class acoustic dataset for drone detection, recorded outdoors with three drone models flying at ranges from about five metres out to roughly two kilometres, and annotated by hand from a single continuous 3.12-hour session. 4 491 drone segments and 1 000 background segments, each 1.0 s, 22 050 Hz, mono, 16-bit PCM. Released with the grouping and split manifests needed for a leakage-free evaluation. Read this before you split the data Segments overlap by 50 %. They were cut from the annotated intervals with a half-second hop, so two consecutive files share half of their samples. A random train/test split therefore evaluates a model partly on audio it has already seen. groups.json maps every segment to the source interval it was cut from (35 annotated drone intervals plus contiguous background blocks). Split at group boundaries, never at file level. Two ready-made protocols are provided in splits/: one interval-disjoint, one temporal (train on the first 65 % of the session, test on the last 20 %). Annotation Drone activity was annotated by hand as time intervals in Audacity, reading the waveform and the spectrogram together. The session was also filmed, and every one of the 35 intervals was checked against the footage; in 21 of them the drone is beyond the frame and the sound itself is the evidence. Background segments come from the opening 25 minutes of the session, in which no drone is heard. Why this dataset exists Four architectures on 128-band log-Mel spectrograms, seven seeds each, using the interval-disjoint split: Training data Balanced accuracy on this set Public drone datasets only 57.5 – 66.0 % Public datasets + this set 98.8 – 99.9 % Models trained on the public acoustic drone corpora do not transfer to these recording conditions, and a quarter of this set's training pool already brings them to 98.2 %. Version 1.1.0 updates the documentation; the audio, groups.json and the split manifests are identical to 1.0.0. Collected and released by the ARIS Laboratory, Department of Electronics and Communication Engineering, Istanbul Technical University.