This page is a work in progress for automated detection of motorboats from underwater soundscape recordings. Some code is messy/incomplete but the workflow below can be executed:
Step by step
- Create audio files, sample-more-audio.ipynb was used for this. This will name files must be named in a certain way. Before the first period '.' it will prefix the filename with: 'y_' if they contain a boat, 'n_' if they do not contain boat noise, then three characters for the site, then the files number (counting up), with underscore to separate each component, e.g n_CCC_106.20220628_144100. May need to check its compatible with your existing filename though. Files can be of any length and must be sampled at 16khz. To simulate real world data there should be many more boat files than non-boat files. Files will be cut into 0.96khz by vggish so the entirety of the file must contain boat noise or none at all.
- Create a conda env using the yml file (vggish-env.yml). This is needed for running feature extraction using the pretrained vggish model.
- Run feature extraction using the pretrained vggish model (Boat-noise-pretrained-CNN-feature-extraction.ipynb). This will require the Audioset folder saved in my other github repo (https://github.com/BenUCL/Reef-acoustics-and-AI). This will output a csv that contains rows for every 0.96s section of each file and the 128 features output by vggish.
- Train an XGBoost classified on the csv (Boat-noise-xgboost.ipynb). This will perform a hyperparameter gridsearch and use 'leave one out cross validation' to maximise the training data and give a highly representative accuracy score. This will then train a final model, save this, and this can be loaded and used on new data.