Dataset of Gunshot Sounds and Koogu Model related to "Impacts of logging, hunting, and conservation on vocalizing biodiversity in Gabon" by Yoh et al. 2024

Description This repository contains gunshot sound training data, testing data, and the Convolutional Neural Network (CNN) model used in our study on gunhunting patterns in Gabon. The dataset includes hundreds of audio recordings containing gunshots and other environmental sounds, their associated annotations, and the trained Koogu machine-learning model. Data Summary Data Types: Gunshot sound recordings, non-gunshot sound recordings, annotations, model files. Data are in the format required for Koogu, where training and test data are contained in folders of audio files with associated folders containing the Raven Pro selection tables. Data Format: Audio files: .wav Annotations: .txt Model: TensorFlow/Keras model (.h5) Code: .py Data Details Training and Test Data: train_annotations "Training_Gunshots_Batch1.txt" is a Raven Pro selection table that contains 203 gunshot annotations that were detected through the spectral cross-correlation analysis. "Training_Gunshots_Batch2.txt" is a Raven Pro selection table that contains 223 gunshot annotations that were detected through the initial Koogu model that was run over the whole dataset. "Training_NotGunshots.txt" is a Raven Pro selection table that contains 8,614 non-gunshot annotations, including sounds from branch snaps, tree falls, calls of the putty-nosed monkey, ambient noise, etc. train_audio "audio_files" contains 184 audio files associated with the selection table "Training_Gunshots_Batch1.txt" "audio_files_second_batch" contains 203 clips associated with the selection table "Training_Gunshots_Batch2.txt". "other_clips" contains 9,274 clips associated with the selection table "Training_NotGunshots.txt". test_annotations "allgunshots_test_gunshots_adjusted_final_clean_20221120.txt" is a Raven Pro selection table that contains the 138 gunshot annotations associated with the files in the "test_audio" folder. This comprises the test dataset. test_audio Contains the 131 recordings associated with the 138 gunshot annotations in "allgunshots_test_gunshots_adjusted_final_clean_20221120.txt" Koogu Model: qDN_4x8_x4_16_Allsites_V3 This is the Koogu model used in this study. It can be run using the script "Koogu_Run_Local_Python_Gabon_ModelV3.py" that you modify depending on the locations of your data and model. CoLab: An example Koogu CoLab page that contains the code used to train and test the model is available here. Data Collection Collection Method: BAR-LT audio recorders deployed in 110 sites within Gabonese national parks, logging concessions, and community forests. Time Period: Recordings from February 2021 to June 2022. Geographical Information: Specific coordinates provided upon request. Data Preparation Downsampling: Recordings were downsampled from 44.1 kHz to 4 kHz. Segmentation: Audio recordings split into consecutive 2.25-second segments with 1.5 seconds of overlap. Normalization: Waveform normalized to [-1.0, 1.0]. Spectrogram Computation: Using a 64 ms window length and 50% overlap, trimmed to 10–1200 Hz frequency range. Training Inputs: Consisted of 584 positive class (gunshots) and 25842 negative class spectrograms. Requirements Software Requirements: Python 3.8, TensorFlow 2.5, NumPy, Pandas. Hardware Requirements: GPU with at least 8GB VRAM recommended. References For more information about this dataset and the model specifications, please consult "Impacts of logging, hunting, and conservation on vocalizing biodiversity in Gabon" (Yoh et al. 2024) How to Cite this Dataset: If you use this dataset in your work, please cite it: Gottesman, B. (2024). Dataset of Gunshot Sounds and Koogu Model related to "Impacts of logging, hunting, and conservation on vocalizing biodiversity in Gabon" by Yoh et al. 2024 (1.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.11192704