https://doi.org/10.5061/dryad.d51c5b0c7

Description of the data and file structure

We present a dataset used to compare traditional bird survey approaches using point counts to the use of autonomous recording units and citizen scientists data at two sites within the Mt Kenya ecosystem.The dataset contains over 20 hours of new recordings obtained from the Mt Kenya ecosystem and annotated by expert ornithologists. The two sites are the Dedan Kimathi University Wildlife Conservancy (DeKUWC) and the Mt Kenya National Park (MKNP).

Files and variables

File: ndege-zetu.zip

Description:

This file contains three folders namely:

• annotations
• audio
• embeddings

The files in these directories are as shown below

├── annotations
│   ├── dekuwc-aru-2016.csv
│   ├── dekuwc-aru-2017.csv
│   ├── dekuwc-kbm.csv
│   ├── dekuwc-pc-2017.csv
│   ├── forest-birds-KE-UG.csv
│   ├── Kenya-Species-List.csv
│   ├── mknp-aru-2017-2018.csv
│   ├── mknp-kbm.csv
│   ├── mknp-pc-2017-2018.csv
│   ├── MKNP-PC-Cues.csv
│   └── single_species_filenames.json
├── audio
│   ├── DeKUWC-10-2017-01-14-06-30-05.mp3
│   ├── DeKUWC-10-2017-01-14-07-00-05.mp3
│   ├── DeKUWC-10-2017-01-14-07-40-05.mp3
...
├── embeddings
│   ├── DeKUWC-1-2016-01-05-12-35-01.npz
│   ├── DeKUWC-1-2016-01-05-16-35-01.npz
│   ├── DeKUWC-1-2016-01-05-16-40-01.npz
...

Annotations

The files in these folder contain

• Annotations of recordings in the audio directory: dekuwc-aru-2016.csv, dekuwc-aru-2017.csv and mknp-aru-2017-2018.csv
• Point count data from the Dedan Kimathi University Wildlife Conservancy (DeKUWC) and the Mt Kenya National Park (MKNP): dekuwc-pc-2017.csv, mknp-pc-2017-2018.csv, MKNP-PC-Cues.csv
• Data from the Kenya Bird Map (KBM): dekuwc-kbm.csv, mknp-kbm.csv
• Audio filenames used to train single species classifiers: single_species_filenames.json
• Species lists: Kenyan bird species (Kenya-Species-List.csv) and forest birds from Kenya and Uganda (forest-birds-KE-UG.csv)

Audio annotations

The audio annotations are weak annotations indicating the species observed in the recording. In addition we indicate whether the species was judged to be in the foreground or background of the recording. Optional remarks are included. Each annotation file has four columns namely:

• Filename
• Foreground Species
• Background Species
• Remarks

Point Count Data

These data are stored as presence absence matrices with the rows representing the species observed and the columns representing the point count.

Kenya Bird Map Data

This is data derived from full-protocal cards and is stored in the same format as the point count data as presence absence matrices with the rows representing the species observed and the columns representing the full protocal card number.

Audio

These are MP3 recordings obtained from DeKUWC and MKNP. The naming convention is {site}-{PC-Location}-YYYY-MM-DD-HH-MM-SS.mp3. Where site will be DeKUWC or MKNP and the PC location is an integer from 1 to 10 representing the 10 point count locations per site. The time the audio recording was captured is encoded in the filename as YYYY-MM-DD-HH-MM-SS

There are 3893 minute long recordings with 1192 of these being new recordings (~20 hours). 2701 are from an earlier dataset

wa Maina, Ciira; Muchiri, David; Njoroge, Peter (2017). Data from: A bioacoustic record of a conservancy in the Mount Kenya ecosystem [Dataset]. Dryad. https://doi.org/10.5061/dryad.69g60

Embeddings

For each recording in the audio directory, we derive dimensional embeddings from the Google Bird Vocalization Classifier (aka Perch). Each minute long recording is divided in to 12 five second segments and a 1280 dimensional embedding is derived for each segment. The 1280 by 12 dimensional array is stored as a numpy array (.npz)

Code/software

This code is used to analyse the data in this repository which compares traditional bird survey approaches using point counts to the use of autonomous recording units and citizen scientists data at two sites within the Mt Kenya ecosystem.

We demonstrate the use of embeddings obtained from Google’s Bird Vocalization Classifier (aka Perch) to train classifiers for the species observed.

Code Directory

The code directory contains the following files.

• bvc_helper_funcs.py : Functions to compute embeddings using the Bird Vocalization Classifier
• birdnet_eval.py: Code to evaluate BirdNET on the annotated data
• bvc_eval.py: Code to evaluate Perch (Google’s Bird Vocalization Classifier) on the annotated data
• birdnet-perch-comp.ipynb: Notebook to compare BirdNET and Perch on annotated data
• compute_file_embeddings.py: A file to compute embeddings from audio files
• mlp-multi-species.ipynb: A notebook to train the multi label multilayer perceptron classifier for Mt Kenya data
• few-shot.ipynb - A notebook describing few shot learning experiments
• forest-species.ipynb - An analysis of forest species observed via point counts, citizen science and autonomous recording units
• lr-mlp-single-species.ipynb - Code to train single species classification models using embeddings
• pc-audio-kbm-cmp.ipynb - Code for statistical comparison of point counts, citizen science and autonomous recording units
• single-species-files.ipynb - Code to detemine audio files with single species
• species-accumulation.ipynb - Code to compute species accumulation curves for point counts
• venn-species-counts.ipynb - Code to produce venn diagrams to analyse species overlap for point counts, citizen science and autonomous recording units. The code also analyses species more often heard than seen in MKNP by processing MKNP-PC-Cues.csv.

To run these files, install the requirements using the procedure described below. An accompanying GitHub repo is https://github.com/DeKUT-DSAIL/ndege-zetu.

Requirements

See the requirements.txt file

Installation procedure

Create a virtual environment

python3.10 -m venv bird-env

The update pip

pip install --upgrade pip

Install the requirements

pip install -r requirements.txt

Access information

2701 of the audio files were was derived from the following sources:

• https://datadryad.org/stash/dataset/doi:10.5061/dryad.69g60 - licensed under a CC0 1.0 Universal (CC0 1.0) Public Domain Dedication license.