Drones on the Danube: where the debris fell, when Romania's phones rang, and what the seismometers recorded

How long after the Izmail siren does Tulcea get an RO-ALERT?#

Izmail is a Ukrainian port on the Chilia arm of the Danube, 19 km from the Romanian town of Tulcea. The Romanian defence ministry's releases describe drone attacks by the Russian Federation on Ukraine's Danube ports since the summer of 2023. When drones head for the region, Ukraine sounds the air-raid alert in Izmail raion (district).

The files of IGSU, Romania's emergency inspectorate, published on data.gov.ro, hold 66 RO-ALERT episodes for an "air attack" between 25 February and 31 August 2026, nearly all in Tulcea or Galați counties. An episode is the first message after a gap of at least three hours. Each episode is set here beside Ukraine's official alerts for Izmail raion (the Ukrainian file stops on 28 August, so the comparison covers 64 episodes).

• 56 of 64 began while Izmail raion was already under alert. Half came at most 31 minutes after the siren; a quarter, more than an hour after.
• Eight came with no alert in Izmail. Where the defence ministry's releases explain them, they describe drones that did not come that way: on 24 July, a target detected 20 km east of Sulina and shot down near Padina; on 24 August, a drone that entered "from the Republic of Moldova".
• Izmail raion was under alert 767 times between 1 February and 28 August, 648 hours in all. Only 62 of those alerts had an RO-ALERT during them or in the half-hour after. The other 705, or 92%, did not. Alerts matched by an RO-ALERT lasted longer: 76 minutes at the median, against 33 for the others.

The half-hour is not a delay. The Ukrainian alert covers the whole raion; an RO-ALERT goes out, according to the ministry's releases, once its radars see targets nearing Romanian airspace, and a drone needs time to reach the river. The data show the order of the two clocks. Why most Izmail sirens never reach a Romanian phone is not in them.

The page RO-ALERT, February–July 2026 counts every message by category. Here the unit is the air-attack episode, not the message: one episode can carry several messages on several channels.

Where drone debris fell in Romania, and how far from Ukraine#

Between September 2023 and 2 October 2026, the press releases of Romania's Ministry of National Defence name 50 places where drone debris, a crater or a crashed drone was found or reported. For 16 of them the release says only that debris was reported or might be from a drone; they are hollow circles on the map. At Tătaru only a crater and remains of an air-to-air missile from the interception were found. The same village can recur: Plauru, on the bank facing Izmail, appears 10 times.

• 2023–2025: 20 sites, all on land, in Tulcea, Brăila and Galați counties. All but one lay within 17 km of Ukraine; half within a kilometre. The exception is the Great Island of Brăila, in March 2024, 49 km away.
• 2026: 30 sites in seven counties. On 25 April, drone debris with "a possible explosive charge on board" "in a built-up area of Galați". On 29 May, a drone "crashed onto the roof of a block of flats" in Galați. On 24 July, after an F-16 shot down a drone that had come in from the sea, drone and missile debris fell near Padina, in Buzău county, 110 km from Ukraine. On 24 August, drone parts in the yard of a house in Ciușlea, Vrancea county, 78 km away.
• The coast: since August 2026, fragments on the beaches at Costinești, Tuzla and Capu Midia, and at sea near the shore from Sulina and Gura Portiței down to Costinești. The releases do not say whether they came from the Danube attacks or from the fighting at sea. Another seven fragments were found far offshore, in Romania's exclusive economic zone, 33–90 nautical miles out; they are not on the map.

The map counts only places named in releases. The table on the ministry's home page (updated 21 October 2026) counts more cases of "drones/fragments identified/lifted/neutralised": 14 in 2024, 16 in 2025, 44 in 2026, 81 since the war began. The releases name the place for 8, 6 and 26 of them. The table cannot be matched case by case with the releases, so the gap cannot be explained from here.

Did the Tulcea and Chilia seismometers record the drone attacks?#

Romania's National Institute for Earth Physics (INFP) runs seismic stations for the Vrancea earthquakes, including two on the Danube: at Tulcea, 17 km from Izmail port, and at Chilia Veche, 37 km away. INFP has already published war explosions from Ukraine picked up by its network (the Khmelnytskyi ammunition depot, May 2023). No public analysis of the stations beside Izmail appears in INFP's publications or the press.

The test was fixed before the 2026 recordings were opened. For each RO-ALERT episode, a window from 15 minutes before the alert to three hours after (three and a quarter hours). For comparison, the same clock hours a week before and a week after, on days without an alert. A "paired bang" is a sudden spike (over ten times the window's usual noise, in the 2–15 Hz band) at Tulcea, with another at Chilia within 150 seconds either way.

• In the 66 alert windows: on average 4.3 paired bangs. In the 89 control windows: 0.5. A statistical test (Mann–Whitney) gives p = 0.001: if alert nights were no different from ordinary ones, a gap at least this large would turn up about once in 1,000 tries.
• On busy nights some pairs form by coincidence. Subtracting the pairs that still form when the Chilia record is shifted by 30 or 60 minutes leaves 2.5 pairs beyond chance in alert windows and 0.1 in control windows. 12 of the 66 alert windows have at least five pairs beyond chance; of the control windows, two.
• Most pairs reach the two stations in the same second; some reach Chilia up to a minute later. That would fit a wave through the ground followed by sound through the air, but it remains a hypothesis: no single bang has been located, and that needs a third station.

The finding is limited to 2026. The same test, run unchanged on 20 attack nights from 2023–2025 dated from the ministry's releases, points the same way (on average 1.6 pairs beyond chance, against 0.3), but the gap is not large enough to rule out chance. A quiet night on the seismometer is not a night without an attack: wind, the size of a blast and distance change what reaches the station. On 29 May, the night of the Galați block, the Tulcea and Chilia stations, 66 and 97 km from Galați, recorded only two pairs. INFP also has a station in Galați; it is not part of the test.

The page shows only the counts and the time of each pair. The recordings belong to INFP (network RO, DOI 10.7914/SN/RO); their reuse terms for waveforms are not published, so the waveforms are not shown here.

What these data cannot tell#

• The list is not complete. It lists the places named in releases. The ministry counts more cases than it names.
• Places are approximate. A dot sits at the village named in the release, not at the exact spot. 15 sites are placed from the release's own description (for instance "about 5 km west of the town of Reni"). Distance to Ukraine is measured from that dot.
• Attribution. The page uses only the ministry's words: attacks by "the Russian Federation" on Ukrainian ports, drones "of the Geran 1/2 type", "of Russian origin", where a release says so.
• Sirens. Why most Izmail alerts do not lead to an RO-ALERT (the drones' route, what the radars see, the alert threshold) is not in the files. The comparison stops on 28 August because the Ukrainian file has no data from 30 August to 1 September.
• Seismometers. A bang cannot be pinned on a particular drone, missile or interception, nor placed on a map. On 19 June, a control window with no alert, no siren and no storm shows 13 pairs beyond chance; it is unexplained and stays in the count.