This repository contains two Jupyter notebooks and the datasets that are described in the arxiv paper HumBug Zooniverse: a crowd-sourced acoustic mosquito dataset.
The notebooks
folder contains two files: metadata.ipynb
contains instructions on how to process and visualise the full dataset with labels found in labels/coarse_data_2sec.csv
, with a breakdown of the sources of recordings and the number of labels obtained from the crowdsourcing on Zooniverse. baseline.ipynb
details an example use, where labels are aggregated from the 2 second overlapping audio recordings, to form a dataset audio_1sec
found in data
with its corresponding label audio_1sec.csv
in labels
. The baseline is a simple convolutional neural network performing classifications on the log-mel feature space with librosa
. We supply a simple cross-entropy weighting for taking into account class imbalance.
We strongly recommend using the processed audio in non-overlapping 1 second segments, as the aggregation has been performed for the four recording groups. The votes supplied in audio_1sec.csv
are given in the categories {yes, no, not_sure}
.
The data is available to download at http://humbug.ac.uk/public/Zooniverse_audio_1sec.zip for the 1 second segments, and at http://humbug.ac.uk/public/Zooniverse_audio_2sec.zip for the overlapping original data. The wave files should be extracted to create the paths: ZooniverseData/data/audio_1sec
and ZooniverseData/data/Zoo_segment
.
The required packages for reproducing the code are given in the first cell of each notebook. This code has been tested in:
Windows 10
Anaconda3 5.2.0 Python 3.6
keras-gpu 2.3.1
installed via conda install -c anaconda keras-gpu