PNW-Cnet: An evolving convolutional neural network to support broad-scale passive acoustic monitoring
| Authors: | Damon B. Lesmeister, Zachary J. Ruff, Adam Duarte, Anna B. Kohlberg, Taal Levi, Christopher M. Sullivan |
| Year: | 2026 |
| Type: | Scientific Journal |
| Station: | Pacific Northwest Research Station |
| DOI: | https://doi.org/10.1016/j.ecoinf.2026.103657 |
| Source: | Ecological Informatics. 94: 103657. |
Abstract
Passive acoustic monitoring is increasingly used for broad-scale, noninvasive tracking of biodiversity across space and time, but processing the resulting audio data at scale remains a major challenge. To address this, we have iteratively developed PNW-Cnet, a convolutional neural network trained to detect wildlife vocalizations in unprocessed field recordings collected throughout a large, forested region of the western United States. Initially built to identify six owl species, PNW-Cnet has evolved through successive retrainings and dataset expansion. We trained PNW-Cnet v5 on 824,120 labeled spectrograms to detect 135 sonotypes (hereafter, classes), including birds (72 species), mammals (11 species), and one amphibian species, as well as diverse anthropogenic and environmental sounds. We also implemented a targeted data augmentation strategy to address class imbalance, enabling the inclusion of rare or underrepresented species and noise types in model training. Here we describe the architecture and training of PNW-Cnet v5, its integration with a user-friendly Python package (pycnet-audio) and a Shiny-based graphical interface for streamlined data processing, and its performance across a range of classes. At a 0.95 confidence score detection threshold, the model achieved ≥95% precision for 112 target classes in real-world applications. Precision remained high across lower confidence score thresholds, while recall varied with class representation in the training set and with vocalization characteristics. PNW-Cnet exemplifies how deep learning models can be recursively expanded and operationalized to support real-world endangered species population monitoring, acoustic community characterization, disturbance mapping, and ecological monitoring. This approach enables the rapid detection of target classes within audio recordings, unlocking new potential for passive acoustic monitoring in biodiversity science and conservation management.