Data from "PNW-Cnet: An evolving convolutional neural network to support broad-scale passive acoustic monitoring"

This repository contains data and code sufficient to reproduce the analyses presented in the associated paper (D. B. Lesmeister et al., "PNW-Cnet: An evolving convolutional neural network to support broad-scale passive acoustic monitoring", in review). For convenience we have grouped the contents into three compressed files: one for each of the two test datasets ("Test_Set_1.zip" and "Test_Set_2.zip") and a third ("v5_Repo_Files.zip") containing the remaining data and code. To completely reproduce the results, you will need to download all three of these compressed files, extract the v5_Repo_Files.zip file to create the required directory structure, and then extract Test_Set_1.zip and Test_Set_2.zip into the corresponding folders in the "data" subfolder. Two different test sets were used to characterize the performance of the PNW-Cnet v5 model. The first, held-out test set ("Test Set 1") comprised 41,202 spectrograms randomly selected from the original training dataset, which were not used for model training or validation. Each spectrogram was generated from a 12-s audio segment and covers a frequency range of 0 to 4000 Hz. These spectrograms were comprehensively labeled with respect to the 135 target classes detected by PNW-Cnet v5. We used this test dataset to calculate class-specific precision, recall, F1 score, and accuracy for the trained model at score thresholds ranging from 0.01 to 0.99. The second, field-based test set ("Test Set 2") was drawn from audio data collected by the Northwest Forest Plan passive acoustic monitoring program at ca. 1000 field sites in western Washington, California, and Oregon, USA, in 2023. This test set includes short audio clips as well as associated spectrograms. They are organized into 85 target classes, each of which produced at least 100 detections at a score threshold of 0.80. We randomly selected 100 apparent detections for each class (total n=8,500 images) and reviewed these spectrograms and the corresponding audio, labeling each image as either a true positive or false positive for the focal class, and used these labels to estimate precision at threshold of 0.80, 0.825, ..., 0.975, 1.00. Site identifiers in both test datasets have been obfuscated so as not to publicize information that could be used to locate breeding territories of northern spotted owls or other species of conservation concern. Additionally, we have omitted audio clips for the "Human" target class in Test Set 2 to avoid potential privacy issues. The README.md file contains detailed information on the output files produced by the different scripts.