The Effect of Soundscape Composition on Bird Vocalization Classification in a Citizen Science Biodiversity Monitoring Project

This archive includes sound clips (.wav files) and associated mel-scale spectrograms of bird vocalizations for 54 species in Sonoma County, California, USA. These data were used for training and validating convolutional neural network (CNN) models for bird species detection. We also include xeno-canto training and validation mel spectrograms used to pretrain CNNs. Details on these data are explained in the paper by Clark et al. (2023) titled "The effect of soundscape composition on bird vocalization classification in a citizen science biodiversity monitoring project". These data are available for use without restrictions, with no warranty on data quality or utility for a given application. We request that any work that does use these data cite the Clark et al. (2023) paper.

Clark, M.L., Salas, L., Baligar, S., Quinn, C., Snyder, R.L., Leland, D., Schackwitz, W., Goetz, S.J., Newsam, S. (2023). The effect of soundscape composition on bird vocalization classification in a citizen science biodiversity monitoring project. Ecological Informatics. https://doi.org/10.1016/j.ecoinf.2023.102065

Associated code for training CNN models, performing inference, and applying post-classification corrections can be found in the GitHub archive https://github.com/pointblue/Soundscapes2Landscapes/tree/master/CNN_Bird_Species

Raw sound data from the Soundscapes to Landscapes project are available upon request: Dr. Matthew Clark, matthew.clark@sonoma.edu

These data were collected as part of the Soundscapes to Landscapes project (soundscapes2landscapes.org), funded by NASA’s Citizen Science for Earth Systems Program (CSESP) 16-CSESP 2016-0009 under cooperative agreement 80NSSC18M0107.

----------------------------
This depository includes the following archives:

• mel_specs.zip: contains 2-sec mel spectrograms split into training (“tr”), validation (“val”), testing (“test”) data for each target bird species (n = 54) used to fine-tune the CNNs. Select spectrogram files are appended with “aug” if they are augmented versions for the training data.

• wav.zip: contains the associated wav-format sound recordings used to generate the training, validation, testing mel spectrograms found in mel_specs.zip.

• Xeno-canto_pretrain.tar: contains 2-sec mel spectrograms split into training and validation data for 40 bird species used for CNN pre-training that were generated using a warbleR segmentation methodology described in the paper. The sound files used to generate these mel spectrograms came from the Kaggle competition, https://www.kaggle.com/datasets/imoore/xenocanto-bird-recordings-dataset
  Mel spectrogram naming reflects the XC number used for cataloging on Xeno-canto in the format XC123456_2.png. The six numbers following the XC characters can be used to search for unique recordings on Xeno-canto (https://xeno-canto.org/) using the search query “nr:123456” in the search tool or queried using the Xeno-canto API (https://xeno-canto.org/explore/api). Unique recording names can be extracted from the mel spectrogram filenames.

• soundscape_test_wavs.zip: the wav-format sound recordings used to perform soundscape testing.