birdidpv identification report

Machine identification, not a verified record: BirdNET scores each 3-second frame. It does not confirm that a species was present. Confidence is not a probability. A high score on one frame is not a record. Listen before citing any of this.

Sources

260115_002_Tr1.WAV (track 1 only)

Recordist

Gewenxin Yu

Recording setup

Two Primo EM273 omni-directional microphones (−37 dB ± 3 dB at 1 kHz, 80 dB S/N ratio, 60 Hz ~ 20 kHz) and a Zoom F3 recorder

Model

BirdNET (Cornell Lab / Chemnitz UT) via birdnetlib, 3-second frames

Ranges analysed

135 below-threshold ranges from lowdom · 00:43:30 of 03:02:24 (23.8% of the take)

Species filter

29.898°N, 102.03°E · week 3 of 48 · 6,522 species reduced to 788

Minimum confidence

0.25

Detections

128 frames in 11 species

The pale band marks the 135 ranges lowdom left below its threshold. The rest of the take was not analysed. Each mark is one 3-second detection. Higher confidence makes the mark taller and darker.

Species

Reference	Species	Frames	Highest	At	First heard
Pavel Shukov · CC-BY-NC	Red-billed ChoughPyrrhocorax pyrrhocorax	82	1.000	00:40:35	00:23:36
egorbirder · CC-BY	Ruddy ShelduckTadorna ferrugineacheck by ear	2	0.928	01:41:05	01:41:05
Joe Bourget · CC-BY-NC	Large-billed CrowCorvus macrorhynchos	17	0.912	00:32:28	00:32:10
Tatyana Zarubo · CC-BY-NC	Great BitternBotaurus stellarischeck by ear	12	0.906	00:11:19	00:10:15
Анна Голубева · CC-BY-NC-ND	Eurasian CurlewNumenius arquatacheck by ear	1	0.684	02:52:35	02:52:35
Frans Vandewalle · CC-BY-NC	Graylag GooseAnser ansercheck by ear	2	0.628	00:09:33	00:09:33
Wei Li Jiang · CC-BY-NC	Eurasian CootFulica atracheck by ear	8	0.611	02:18:36	00:23:15
zametnya · CC-BY-NC	Eurasian WigeonMareca penelopecheck by ear	1	0.443	00:37:06	00:37:06
unknown photographer · CC-BY-SA	MallardAnas platyrhynchoscheck by ear	1	0.439	02:08:17	02:08:17
thegreatdodo · CC-BY-NC	Common ShelduckTadorna tadornacheck by ear	1	0.365	00:24:40	00:24:40
Frank Sengpiel · CC-BY	Gray HeronArdea cinereacheck by ear	1	0.264	02:37:10	02:37:10

Reference photographs come from iNaturalist. We store them in this package rather than hot-linking them. The package survives archiving that way, and no reader's address is sent to a third party. Each file is kept exactly as iNaturalist served it. It remains under its photographer's licence, credited beside it. The machine-readable record is in credits.json. A photograph shows what the species looks like. It is not evidence that the species was here. Nine of these eleven rows are ones we ask you to doubt. One of those nine is a photograph of a bird that was never there.

Taxon identifiers

These are the same eleven species in the databases a detection usually has to travel to. We matched them on iNaturalist taxon ID, not on name. A name search returns a congener often enough. Attaching an identifier to the wrong bird is a real risk. Wikidata holds dozens more per species. We keep those in taxon-ids.json.

Reading this against a frozen lake

Two corvids account for 99 of the 128 detections. They behave like real calls. They carry energy well above 2 kHz. They sit 3–11 dB above everything else. The other nine species are all deep-voiced waterbirds. The lake had frozen over by mid-January at 4,055 m. We measured the detected frames after a 150 Hz high-pass. That measurement shows where the two groups part:

The waterbird detections concentrate their energy below 400 Hz at a lower level. They mostly show none of the high-frequency structure the corvid detections do. That band is where the ice resonance sits. Local guides call that sound long hou, “dragon roar”. A plausible reading is that BirdNET maps ice onto birds whose calls are booms. Great Bittern at 0.906 is the clearest case. Its twelve detections arrive in sustained runs. Eight of them fall inside seventy seconds. Not one of their spectrograms shows a call above the ice.

Listening to the species explains the mistake. It does not excuse it. On the first page of xeno-canto recordings for Botaurus stellaris, try XC891071, XC1000766, XC832807 and XC741523. The booms really are close to the dragon roar. The confusion looks reasonable rather than absurd. Listen to them here. Then listen to any green stretch of the timeline above:

• XC891071 · player · Christian Bøggild · Denmark · CC BY-NC-SA
• XC1000766 · player · Christian Bøggild · Denmark · CC BY-NC-SA
• XC832807 · player · Cedric Mroczko · Ukraine · CC BY-NC-SA
• XC741523 · player · Romuald Mikusek · Poland · CC BY-NC-SA

These recordings are stored in this package under their recordists' CC BY-NC-SA licences. We do not embed them from xeno-canto. The comparison survives archiving that way, and no reader's address is sent to a third party. We re-render the spectrograms here with the same axes and scaling as the detection frames above. xeno-canto draws its own spectrograms on a linear frequency axis. That axis suits the wide scrubbing strip in its player. It also leaves the boom as a thin line along the bottom while the background birds fill the frame. Put next to a detection from this take, the same sound would look like a different one.

One thing separates them by ear. The ice carries an electronic quality that the bird does not. That difference is the interesting part of the error. It is plainly audible. None of the spectrogram statistics on this page found it.

Ice is not the only thing here that is not a bird. The frame at 01:41:05 is a person. BirdNET calls it Ruddy Shelduck at 0.928. That is the second highest confidence in the whole run. It is Xiao Zhang, one of the guides. He was calling across the lake to the recordist. The recordist identified it on listening. There is no bird in those three seconds at all.

That frame is worth opening. It also shows how a spectrogram can be misread. It carries harmonic stacks between 1 and 3 kHz. We first took them for a call. Ice does not produce structure like that. We moved the species out of the doubtful list. A voice does produce it. The stacks are a pitch contour gliding through its harmonics. Formants sit around 1.2 and 2 kHz. Four or five syllables appear in the last second. That is speech, not song. The band average in the table above had hidden the voice under the low-frequency bed. The picture that uncovered it was then read wrong. Only listening settled it.

The ledger runs three ways, not two. It includes real calls, lake ice, and at least one human voice. Every row marked check by ear is one we would not cite without going back to the source WAV. The loudest argument for that rule is simple. The model's second most confident bird in three hours was a man shouting.

Open the multimedia report Open the bare CLI multimedia report Open the bare CLI timeline report Open the lowdom screening report Download every detection Download the species roll-up Download the taxon identifiers Read the summary

Identification by BirdNET, developed by the K. Lisa Yang Center for Conservation Bioacoustics at the Cornell Lab of Ornithology with Chemnitz University of Technology, reached through birdnetlib. The BirdNET models are licensed CC BY-NC-SA 4.0 and that non-commercial condition applies to these results. Cite: Kahl, S., Wood, C. M., Eibl, M., & Klinck, H. (2021). BirdNET: A deep learning solution for avian diversity monitoring. Ecological Informatics, 61, 101236. Neither this report nor field-audio-tools is affiliated with or endorsed by the BirdNET team.