Soundscape Connectomes: Unsupervised Graph-Based Approach for Soundscape Mapping

We introduce soundscape connectomes, which are graph representations of the acoustic relationships within a landscape, where nodes represent geographical sites and edges reflect relations derived from each site’s biophony. Soundscape connectomes are constructed from passive acoustic monitoring (PAM) recordings, enabled by the unique acoustic signatures of habitats. However, in ecoacoustic analysis, ground-truth graphs or labels are often not available. We propose an unsupervised pipeline that decomposes recordings into sonotypes, builds per-site acoustic structures, infers graphs with several methods, and compares them using a smoothness-based, unsupervised criterion that scores reconstruction of held-out nodes. We apply the proposed method to a large-scale real-life data set acquired in the Colombian Andes, comprising over 19.598 recordings from 17 sites and 290 sonotypes. Results show that different graph inference methods yield comparable connectome structures, supporting a practical criterion for selecting graph models without prior information. This approach positions soundscape connectomes as a complement to remote-sensing analyses for ecosystem monitoring and conservation.