Acoustic drone detection at the front
On 14 August 2026 a preprint by Vadym Vilhurin, Volodymyr Sydorskyi and Andrii Shevtsov appeared on arXiv on passive acoustic detection of small drones in battlefield conditions. On a set of recordings from Ukrainian frontline zones the authors raised the F1 score from 55.4% to 78.6%.
Why it matters
An example of applied Ukrainian military research published openly: a cheap passive sensor trained on recordings from the front rather than from a test range. The result concerns one dataset supplied by the company Zvook; the text does not say the set is open, so the figures cannot be checked independently.
What the paper says. The authors propose per-channel energy normalisation (PCEN) and attention-based pooling, and a domain-aware training strategy with auxiliary classes and data from several microphones. The backbone is ConvNeXt-Tiny. The data come from two microphone types: Mic-1 on communication towers at heights from 10 to 40 m, Mic-2 about 1 m above ground at the front; in the small training set Mic-2 gives under 1% of samples and the small-drone class has 441 samples. The test set was collected in spring "strictly from active combat zones". The abstract: F1 from 55.4% to 78.6%. At the end the paper gives 55.4% for the open-source AST-Drone model and 82.2% for the authors' ensemble on the full training set. The work was done at the Institute for Applied System Analysis of Igor Sikorsky Kyiv Polytechnic Institute (Vilhurin's master's research under Sydorskyi); the authors are also listed with Zvook (Lviv) and the Kyiv School of Economics. What the record does not claim: that the system is used by the military; that the dataset is open; results on other data. All figures are the authors' own measurements. Version v1, the only one at the time of writing, was read.