Bioacoustic Geolocation: Species Sounds as Geographic Signals

Technical Summary: Bioacoustic Geolocation: Species Sounds as Geographic Signals

Problem Definition
This work addresses the challenge of global-scale audio geolocation: determining the geographic coordinates of an audio recording solely based on its acoustic content. Unlike image geolocation, which relies on salient visual cues like landmarks or architectural styles, audio lacks explicit, globally consistent geographic markers. The authors hypothesize that bioacoustic signals—specifically the sounds of wildlife—contain informative geolocation cues due to the well-defined geographic ranges of species. The core intuition is that identifying the species present in a recording allows for the intersection of their respective range maps, thereby constraining the possible recording location.

Methodology
The study utilizes the iNatSounds dataset (230k recordings, 5,500+ species) and introduces XCDC, a new benchmark of species-rich dawn chorus recordings curated from Xeno-Canto. The authors evaluate several approaches:

1. Baseline Methods:

   • Regression: Directly predicting latitude and longitude using Euclidean or Haversine distance losses.
   • Classification: Predicting geographic bins using hexagonal grids (H3) at varying resolutions (coarse to fine) and hierarchical approaches.
   • Retrieval: Adapting image geolocation retrieval paradigms (e.g., GeoCLIP, SatCLIP) to audio, where audio features are projected into a location embedding space to find the most similar pre-computed location.

2. Proposed Approach (AG-CLIP):
   The authors propose AudioGeo-CLIP (AG-CLIP), a hybrid retrieval-based method that extends GeoCLIP.

   • Architecture: It employs a vision-inspired audio encoder (MobileNet-V3) that processes spectrograms.
   • Dual-Task Learning: In addition to the standard contrastive loss for aligning audio with location embeddings, AG-CLIP includes an auxiliary task: predicting species checklists (presence/absence of species) derived from the SINR (Spatial Implicit Neural Representations) range map model.
   • Rationale: This auxiliary loss encourages the encoder to attend not only to foreground vocalizations but also to background ambient sounds, learning associations between the acoustic environment and the species likely to inhabit that habitat.
   • Variable Length Handling: Recordings are processed as collections of 3-second clips. Predictions are aggregated (via averaging) to form a final location estimate.

3. Oracles and Diagnostics:
   To establish upper bounds, the authors construct "Species Oracles" using ground-truth species checklists derived from SINR range maps. This isolates the theoretical potential of species information from the difficulty of species identification.

4. Spatiotemporal Aggregation:
   The study investigates improving performance by aggregating multiple short recordings from the same spatiotemporal neighborhood (e.g., same H3 hexagon over a week, month, or year) to simulate a more comprehensive species checklist.

Key Results

• Oracle Performance: When provided with a perfect "full checklist" of species present at a location, the model achieves near-perfect regional accuracy (97.8%), confirming that species range data is a powerful geolocation signal. However, performance degrades significantly with partial or corrupted checklists, highlighting the difficulty of capturing all species in short recordings.
• Model Performance: AG-CLIP outperforms all baselines on the iNatSounds test set, achieving a median geolocation error of 1,082 km.
  • Accuracy: It reaches 17.2% accuracy at the regional scale (200 km) and 41.0% at the country scale (750 km).
  • Comparison: It significantly outperforms off-the-shelf soundscape mapping models (GeoCLAP, Taxabind) and standard regression/classification baselines.
  • Ablations: The auxiliary checklist loss improves continental accuracy by 2.7%. Pretraining on iNatSounds (vs. ImageNet or general audio datasets like CLAP) is critical for performance.
• XCDC and Aggregation:
  • Performance drops on the XCDC dataset (species-rich dawn choruses), suggesting current models struggle with the distribution shift and high species overlap found in complex soundscapes.
  • Spatiotemporal aggregation significantly improves results. Aggregating recordings across a full year within a 36 km² area improves regional accuracy from 17.2% to 25.9%.
• Multimodal Geo-Forensics: The authors demonstrate that AG-CLIP can detect inconsistencies between audio and visual modalities in movie clips (e.g., Dune, Star Wars), where audio geolocation predictions diverge from visual scene predictions, indicating potential post-production mismatches.

Significance and Claims
The paper claims to be the first study on global-scale bioacoustic geolocation. Its primary contributions are:

1. Formalizing the Problem: Establishing a benchmark for audio geolocation that moves beyond urban noise to include wildlife and natural sounds.
2. Validating the Hypothesis: Demonstrating that species range information is a strong, albeit underutilized, signal for geolocation, and that models can implicitly leverage these signals.
3. Methodological Advancement: Proposing AG-CLIP, which effectively combines species range prediction with retrieval-based geolocation, outperforming existing soundscape mapping methods.
4. Practical Insights: Showing that while short recordings limit precision, aggregating data across space and time (a practical approach for citizen science apps like iNaturalist) substantially improves accuracy.

The authors maintain a modest stance, acknowledging that audio geolocation remains significantly more challenging than image geolocation (with regional accuracies of ~17% vs. ~30-60% for images). They attribute this to the acoustic self-similarity of distant regions, the large ranges of many species, and the difficulty of capturing rare but geographically distinctive species in short clips. The work is positioned as a foundational step toward future applications in ecological monitoring, digital forensics, and privacy-preserving audio processing.