Awesome

Nighthawk

Overview

Nighthawk is a machine learning model for acoustic monitoring of nocturnal bird migration.

Nighthawk is trained on recordings of nocturnal flight calls (NFCs) from the Americas, with the greatest coverage in eastern North America. It processes mono-channel audio files in .wav format and returns detections in a tabular format (.csv and .txt). Nighthawk currently includes training data from 82 species, 18 families, and 4 orders of birds that vocalize during nocturnal migration.

Nighthawk runs in a Python environment. The underlying model was trained with TensorFlow.

For details on Nighthawk training and performance, see the following paper:

Van Doren, B. M., A. Farnsworth, K. Stone, D. M. Osterhaus, J. Drucker, and G. Van Horn. 2023. Nighthawk: acoustic monitoring of nocturnal bird migration in the Americas. bioRxiv. doi:10.1101/2023.05.22.541336.

<a rel="license" href="http://creativecommons.org/licenses/by-nc/4.0/"><img alt="Creative Commons License" style="border-width:0" src="https://i.creativecommons.org/l/by-nc/4.0/88x31.png" /></a><br />This work is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-nc/4.0/">Creative Commons Attribution-NonCommercial 4.0 International License</a>.

Installation

We recommend installing and using Nighthawk in its own Python environment. If you're new to Python environments, we suggest installing either Anaconda or Miniconda and using the conda program that comes with them to manage your environments. For help choosing between Anaconda and Miniconda, see here.

The following instructions assume that you're using conda to manage your Python environments. If you're using anything else (such as venv, for example), you'll need to modify the instructions accordingly.

On Windows, type the commands below in an Anaconda Prompt. Anaconda Prompt is a program that comes with Anaconda and Miniconda. It is a lot like the built-in Windows Command Prompt program, but it is customized for Anaconda and Miniconda. Once you've installed Anaconda or Miniconda, you can find the Anaconda Prompt program in your Windows Start menu or by typing "Anaconda Prompt" in the search field of the Windows taskbar. On macOS and Linux, type the commands below in a regular terminal.

To install Nighthawk, first create a new Python environment named nighthawk-0.3.1 that uses Python 3.10:

conda create -n nighthawk-0.3.1 python=3.10

Then activate the environment:

conda activate nighthawk-0.3.1

and install the Nighthawk Python package and its dependencies:

pip install nighthawk

Finally, if (and only if) you're installing on an Apple computer with an Apple silicon (i.e. M1 or M2) processor, issue the following command:

conda list

to list all of the Python packages that are now installed in your environment. If the list includes a package called tensorflow and not a package called tensorflow-macos, then uninstall the tensorflow package with:

pip uninstall tensorflow

and install tensorflow-macos with:

pip install tensorflow-macos

For more about conda, including additional conda commands, see its documentation.

`tensorflow-metal` for Apple Silicon Macs

On Apple computers with Apple silicon (i.e. M1 or M2) processors, installing the tensorflow-metal Python package into your Nighthawk Python environment with:

pip install tensorflow-metal

can accelerate Nighthawk considerably. Unfortunately, however, version 1.1.0 of the tensorflow-metal package (the most recent version as of this writing) also appears to leak memory and cause Nighthawk to hang when processing large (e.g. full-night) audio files on our test machine, an M2 Mac mini with 8 GB of memory. For this reason the package is not installed with Nighthawk by default. Feel free to install it yourself, however, and test it for your particular situation.

Command Line Usage

Once you've installed Nighthawk, the simplest way to run it is to invoke the nighthawk program from the command line. At an Anaconda Prompt (on Windows) or terminal (on macOS or Linux), first activate your Nighthawk environment with:

conda activate nighthawk-0.3.1

Then, to run the nighthawk program on an audio file, say my_file.wav:

nighthawk my_file.wav

nighthawk will output detections to the file my_file_detections.csv in the same directory as the input.

Adding the --raven-output flag will also export a my_file_raven.txt file that can be read as a selection table by Raven Pro:

nighthawk my_file.wav --raven-output

Adding the --audacity-output flag will also export a my_file_audacity.txt file that can be read as a label track by Audacity:

nighthawk my_file.wav --audacity-output

You can also specify more than one input file for nighthawk to process, and specify relative or absolute file paths as well as just file names.

nighthawk has several command line arguments for configuring detection parameters and controlling output. For full nighthawk help, including a list of all command line arguments, run the command:

nighthawk --help

to produce the output:

usage: nighthawk [-h] [--hop-size HOP_SIZE] [--threshold THRESHOLD] [--ap-mask AP_MASK] [--merge-overlaps | --no-merge-overlaps] [--drop-uncertain | --no-drop-uncertain] [--calibration | --no-calibration]
                 [--quiet | --no-quiet] [--csv-output | --no-csv-output] [--raven-output | --no-raven-output] [--audacity-output | --no-audacity-output] [--duration-output | --no-duration-output]
                 [--gzip-output | --no-gzip-output] [--tax-output | --no-tax-output] [--output-dir OUTPUT_DIR_PATH]
                 input_file_path [input_file_path ...]

positional arguments:
  input_file_path       path of audio file(s) on which to run the detector. Accepts directories and wildcards.

options:
  -h, --help            show this help message and exit
  --hop-size HOP_SIZE   the hop size as a percentage of the model input duration, a number in the range (0, 100]. (default: 20)
  --threshold THRESHOLD
                        the detection threshold, a number in [0, 100]. (default: 80)
  --ap-mask AP_MASK     the AP mask threshold, a number in [0, 1]. (default: 0.7)
  --merge-overlaps, --no-merge-overlaps
                        merge overlapping detections. (default: True)
  --drop-uncertain, --no-drop-uncertain
                        apply postprocessing steps to drop less certain detections. (default: True)
  --calibration, --no-calibration
                        calibrate model outputs. (default: True)
  --quiet, --no-quiet   Mask unnecessary console messages. (default: False)
  --csv-output, --no-csv-output
                        output detections to a CSV file. (default: True)
  --raven-output, --no-raven-output
                        output detections to a Raven selection table file. (default: False)
  --audacity-output, --no-audacity-output
                        output detections to an Audacity label file. (default: False)
  --duration-output, --no-duration-output
                        output file duration in seconds to a txt file. (default: False)
  --gzip-output, --no-gzip-output
                        gzip all output files. (default: False)
  --tax-output, --no-tax-output
                        save separate taxonomic output files. (default: False)
  --output-dir OUTPUT_DIR_PATH
                        directory in which to write output files. (default: input file directories)

Use with Vesper

You can also use Nighthawk from Vesper, for example if you would like to view and interact with spectrograms of its detections in Vesper clip albums. For more on this see the Vesper documentation.

Improving Nighthawk by correcting mistakes

See this page for instructions on how to review Nighthawk's results on your data and submit files to us to improve the model.

Licensing and Citation

Under this license, you are free to share and adapt this material, but you must provide appropriate attribution, and you may not use the material for commercial purposes.

To reference Nighthawk, please use the following citation:

BibTex:

@article{vandorenNighthawkAcousticMonitoring2023,
    title = {Nighthawk: acoustic monitoring of nocturnal bird migration in the Americas},
    url = {https://www.biorxiv.org/content/10.1101/2023.05.22.541336},
    doi = {10.1101/2023.05.22.541336},
    journaltitle = {{bioRxiv}},
    author = {Van Doren, Benjamin M and Farnsworth, Andrew and Stone, Kate and Osterhaus, Dylan M and Drucker, Jacob and Van Horn, Grant},
    date = {2023},
}

Code of Conduct

See our code of conduct.

Contact and Collaborations

Please contact Benjamin Van Doren (vandoren@cornell.edu) with questions about Nighthawk. We are always open to scientific collaborations.