Nocturnal flight calls dataset: long-term acoustic monitoring of birds migrating at night

General Description:

This is a development set used in the experiments in the Ph.D. thesis: "Nowe metody akustycznej identyfikacji ptaków migrujących nocą" ("Novel methods of acoustic identification of birds migrating at night") by Hanna Pamula. The project focuses on the detection (and - partially - classification) of passerine birds' calls from long-term audio recordings collected during bird autumn migration between 2016 and 2019. The dataset consists of >56,5 hours of recordings with annotations of nocturnal flight calls of passerine birds migrating along the Baltic Sea coast, Poland.

Folder Structure

Development_Set_3.1.zip

|_Development_Set_3.1/

    |__Training_Set/

            |____*.wav

            |____*.txt

    |__Validation_Set/

            |____*.wav

            |____*.txt

    |__Testing_Set/

            |____*.wav

            |____*.txt

Training Set: 86 recordings

Validation Set: 8 recordings

Testing set: 18 recordings (BUT: uploaded 20 recordings, as in the previous version of the dataset - version 3, two additional recordings were used. Then, they were deleted in the final version of development set 3.1. Two additional recordings are: 'BUK5_20161101_002104a and BUK5_20161101_002104b)

Names of waveforms and annotations are matching.

Waveforms:

The whole dataset consists of 114 recordings. One hundred thirteen recordings are about 30 minutes long (29min56s – 29min 59s), one recording is 1min20s. All data were recorded at 44,100 Hz sampling rate, one channel, with SM2 Wildlife Acoustics recorders + SMX-NFC microphone. The recording sessions were performed at night (starting time and date denoted in a file name) on the Baltic Sea coast in Poland (Dąbkowice, near Darłowo).

Annotations:

Transcriptions were produced using Audacity 2.4.1: https://www.audacityteam.org/ by an experienced birdwatcher, Hanna Pamula. While every effort has been made to ensure the quality and accuracy of the labels, some errors may occur, taking into account the difficulty of nocturnal call recognition and transcription tasks in general.

Transcription format:

[Starting time (sec)] [Ending time (sec)] [Label]

Meaning of the labels:

1. Positive classes – migrating passerine birds:

• 's' – song thrush call (Turdus philomelos)
• 'k' – blackbird call (Turdus merula)
• 'd' – redwing call (Turdus iliacus)
• 'r' – robin call (Erithacus rubecula)
• ‘kwiczol’ – fieldfare call (Turdus pilaris)
• ‘skowronek’ – skylark call (Alauda arvensis)
• Each of the above labels could also have a question mark '?', e.g. 'r?', 'k?' – meaning that it's not a sure label. In a bird call detection task, they are regarded as positive chunks containing bird call(s).
• 'ni' – non identified bird call (distant/quiet/not recognized)

Only the supposed calls of migrating passerine birds were labeled; other sounds of species were ignored (e.g., robin's tik-calling, which can be often heard at dusk, and may be regarded as warning sounds).

2. Negative classes – other marked sound events:

• 'g' – other bird calls/songs/sounds. Sounds that could confuse the model; for example, sounds of migrating geese, cranes, plovers calls, etc.
• 'gh' – human voices
• 't' – cracks, clicks, raindrops, other noise
• ‘puszczyk’ – tawny owl voice (Strix aluco)
• 'czapla' – grey heron voice (Ardea cinerea)

Not all occurrences of the negative sounds were labeled – only some chosen examples to represent the possible noises/negative samples. Thus these annotations can't be used for entirely different detection / classification tasks than intended, e.g., detecting migrating cranes or human voices in long-term recordings.

3. Labels to be excluded from analysis:

• '???', '??? mysz', '??? high freq' – unknown, not sure if the sound event is a birds' call or not. Uncertainty about belonging to a positive/negative class in the detection task.