📊 statistics

Physics-Informed Learning for Robust Acoustic Localization with Calibrated Uncertainty

This paper proposes a physics-informed learning method that refines classical hyperbolic acoustic localization to achieve robust, uncertainty-calibrated performance in complex outdoor environments, thereby enabling scalable automated wildlife monitoring.

Original authors: Jennifer N. Kampe, Changwoo J. Lee, Xin Shen, Ari Lehtiö, Sandro von Brandenburg, Ossi Nokelainen, David B. Dunson, Otso Ovaskainen

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jennifer N. Kampe, Changwoo J. Lee, Xin Shen, Ari Lehtiö, Sandro von Brandenburg, Ossi Nokelainen, David B. Dunson, Otso Ovaskainen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to find a friend in a massive, echoing canyon just by listening to their voice. You have a team of friends standing in a circle, each holding a stopwatch. When your friend shouts, everyone notes exactly when the sound hits their ear. By comparing the tiny differences in arrival times, you can mathematically draw a map to pinpoint where the shout came from. This is the basic idea behind "Passive Acoustic Monitoring," a superpower scientists use to track wildlife like birds, bats, and whales without ever having to see them. It's like being a detective who solves crimes using only sound.

However, the real world is messy. In a canyon, sound doesn't just travel in a straight line; it bounces off cliffs, trees, and the ground. These echoes can trick your friends' stopwatches, making them think the sound came from a different direction. This is called "multipath," and it turns a simple math problem into a confusing puzzle. Traditional methods, which rely on strict physics formulas, often work perfectly in a quiet, empty room but can fall apart completely in a noisy forest, leading to wild guesses about where an animal is hiding. The big question scientists are asking is: How do we keep the speed of math but add the smarts to ignore the confusing echoes?

This paper introduces a clever new way to solve that puzzle by teaching a computer to be a "physics-savvy detective." Instead of throwing away the old math formulas, the authors built a hybrid system. They take the fast, traditional method (the "hyperbolic solver") and give it a safety net. When the math gets confused by echoes or bad weather, a smart computer model steps in to fix the mistake. Think of it like a GPS that knows when it's lost. If the GPS says you are in the middle of a lake (which is impossible), the system checks its own confidence. If it's unsure, it switches to a backup plan that uses clues about the environment to guess a better location.

The researchers tested this idea using a mix of computer simulations and real-world experiments. They set up microphones in two very different places: a frozen, treeless lake and a dense, noisy forest. In the forest, where sound bounces off trees and creates the most confusion, the old math method made huge mistakes, sometimes guessing the animal was 70 meters away when it was actually right next to the microphone. The new hybrid method, however, caught these errors. While the old method had a "median" (middle) error of 0.14 meters in the forest simulation, the new method's median was slightly higher at 0.38 meters. But here is the magic: the new method stopped the "catastrophic" errors. It kept the worst-case guesses down to about 15 meters, whereas the old method let them spiral up to 71 meters.

The paper also teaches the system how to say, "I'm not sure." It doesn't just give a single point on a map; it draws a circle around that point to show how confident it is. If the microphones are arranged in a tricky shape that makes it hard to tell direction, the circle gets bigger. If the geometry is perfect, the circle shrinks. The authors found that by using a special "geometry-aware" adjustment, their confidence circles were much more accurate, covering the true location about 78% of the time when they aimed for 80% confidence, compared to only 73% with the standard method.

Crucially, the authors are careful to note that while the forest results are very promising, they are currently based on simulations. The real-world forest data is still being collected and analyzed. However, the frozen-lake tests, which used real recordings, showed that the system works in the real world, matching the old method's accuracy on easy days while dramatically reducing the chance of a total failure on bad days. The paper suggests that this approach could be a game-changer for monitoring wildlife at a massive scale, allowing scientists to track animals in complex environments with a level of reliability and honesty about uncertainty that previous methods simply couldn't provide.

Technical Summary: Physics-Informed Learning for Robust Acoustic Localization with Calibrated Uncertainty

1. Problem Statement

Passive Acoustic Monitoring (PAM) offers unprecedented opportunities for ecological spatial point-process data collection. However, realizing this potential requires accurate and scalable localization methods. Classical localization techniques, primarily hyperbolic (Time Difference of Arrival, TDOA) and score-based methods, rely on simplifying assumptions that are routinely violated in real-world outdoor soundscapes.

Key challenges include:

Multipath Dominance: Reflected paths often exceed direct paths in amplitude, causing generalized cross-correlation with phase transform (GCC-PHAT) to select incorrect peaks.
Environmental Misspecification: Unobserved or weakly parameterized conditions (e.g., temperature-dependent sound speed, habitat-dependent attenuation) degrade model performance.
Non-Acoustic Interference: Mechanical vibrations or impulsive events bypass the acoustic propagation model, causing hard failures.
Spatial Aliasing: Experimental design limitations (array geometry, bandwidth) can render distinct source locations indistinguishable.

In these regimes, classical methods become "brittle," prone to catastrophic worst-case errors even in small detection arrays. Furthermore, standard deep learning approaches struggle with tabular data due to a lack of inherent structure (unlike images or text) and often fail to provide calibrated uncertainty, which is essential for downstream spatial modeling.

2. Methodology

The authors propose a hybrid approach that refines, rather than replaces, underlying physical laws. The method utilizes a Prior-Data Fitted Network (PFN), specifically TabPFN, operating on a carefully engineered set of physics-informed acoustic features.

2.1 Feature Engineering

Instead of processing raw waveforms, the system summarizes each acoustic event into a row of 72 tabular features. These features are designed to expose physically meaningful cues:

TDOA Features: Multiple GCC peak candidates (not just the dominant one) and cross-correlation confidence scores.
Uncertainty Indicators: Reliability measures derived from GCC structure (e.g., peak separability, lag gaps) to flag ambiguity.
Propagation Features: Impulse response summaries (energy decay, reverberation) and spectral features capturing attenuation and coloration.
Contextual Features: Array geometry descriptors and environmental parameters (temperature, noise level).
Solver Outputs: The closed-form hyperbolic solver's point estimate and fit diagnostics are included as input features.

2.2 Hybrid Point Estimation

The core of the method is a three-component system built atop a fast hyperbolic solver:

Learned Refinement: A TabPFN model is trained to regress source coordinates ( $x, y$ ). It takes the hyperbolic solver's output as an input feature, effectively learning to correct the physics-based solution rather than localizing from scratch.
Adaptive Gating: To prevent the learned model from degrading reliable solutions, a two-stage gate determines when to substitute the solver's estimate with the TabPFN prediction:
- Geometric Check: Rejects estimates falling outside a bounding box expanded from the microphone array.
- Energy-Consistency Check: Computes the RMS residual between observed log-amplitude ratios and those predicted by the candidate position under spherical spreading ( $1/r$ ). If the residual exceeds a threshold $\tau^*$ , the solution is flagged as unreliable.
- Decision: The TabPFN prediction is used only if the hyperbolic solution fails these plausibility checks; otherwise, the solver's estimate is retained. This bounds catastrophic tail errors while preserving median accuracy.

2.3 Calibrated Uncertainty Quantification

The method provides geometry-aware uncertainty estimates suitable for propagation into downstream models:

Error Modeling: A second TabPFN model predicts the localization error of the hybrid estimate using features excluding the hyperbolic point estimate.
Conformal Calibration: To ensure finite-sample coverage guarantees, the predicted error quantiles are calibrated using conformal prediction.
GDOP Scaling: The authors introduce a Geometric Dilution of Precision (GDOP)-scaled correction. Unlike fixed conformal intervals, this approach scales the prediction interval width based on the local array geometry. Intervals widen where the geometry is ill-conditioned, ensuring coverage is maintained in difficult regions without unnecessarily widening intervals in easy ones.

2.4 Training Data Simulation

Training data is generated using ForestIR, a state-of-the-art physics-based simulator. Synthetic source positions are drawn within and slightly outside the array's convex hull. Simulations incorporate:

Real bird vocalizations mixed with site-specific background noise regimes.
Physics-based propagation modeling (direct path, ground reflection, trunk scattering).
Environmental conditions (temperature, humidity) sampled from field campaign ranges.

3. Key Results

The method was evaluated on distributed microphone arrays in both simulated and real outdoor environments.

3.1 Frozen-Lake Site (Real Field Data)

Setup: A 3-microphone array in a low-noise, treeless environment.
Performance: The hybrid gate matched the hyperbolic solver's median error (0.41 m) but significantly improved the upper tail.
- 90th Percentile: Reduced from 24.6 m (solver) to 22.3 m (hybrid).
- 95th Percentile: Reduced from 42.1 m (solver) to 35.3 m (hybrid).
Observation: Simulation errors were significantly lower than field errors (median 0.04 m vs. 0.41 m), highlighting the difficulty of modeling near-surface micro-climatic effects. This underscores the necessity of a short calibration step using real playback data.

3.2 Forest Site (Simulation Only)

Setup: A 6-microphone array in a dense, heterogeneous forest with trunk scattering.
Performance: This environment represents the method's target regime.
- Median Error: The hyperbolic solver was sharper (0.14 m) compared to TabPFN (0.38 m).
- Tail Robustness: The solver's error tail exploded under multipath conditions (95th percentile: 56.6 m; 99th percentile: 71.4 m). In contrast, the TabPFN-based method maintained much tighter bounds (95th percentile: 9.26 m; 99th percentile: 14.8 m).
Conclusion: The hybrid approach substantially reduces catastrophic errors characteristic of hyperbolic failures in complex environments.

3.4 Uncertainty Calibration

Coverage: GDOP-scaled conformal calibration improved empirical coverage at the 0.80 nominal level from 0.73 (fixed) to 0.78, recovering most of the shortfall without additional data.
Discriminative Power: The predicted error was highly discriminative (Spearman $\rho \approx 0.53$ ), ranking high-error detections significantly better than a purely geometric baseline ( $\rho \approx 0.22$ ).

4. Significance and Claims

The paper claims that this approach provides a step toward scalable automated wildlife monitoring in complex acoustic environments. Key contributions include:

Robustness via Hybridization: Rather than statistically replacing physics, the method refines it. By using a learned model to correct a fast hyperbolic solver only where it produces implausible solutions, the system substantially reduces catastrophic worst-case errors while matching median accuracy.
Calibrated Uncertainty: The provision of geometry-aware, calibrated uncertainty estimates allows for the propagation of detection reliability into downstream spatial ecological models, a critical requirement often missing in automated pipelines.
Scalability: The pipeline is designed for scalability. Site characterization (array geometry, tree locations via satellite imagery) and simulation setup are one-time tasks. A short playback experiment from known positions anchors the model and calibrates uncertainty, requiring no sophisticated equipment beyond the monitoring array itself.
Efficiency: By leveraging tabular representations of physics-informed features, the method avoids the computational cost and data hunger of learning directly from raw waveforms, making it feasible for large-scale deployment.

The authors remain modest, noting that while preliminary results are promising, field validation in the forest regime is ongoing, and the gap between simulation and real-world performance necessitates careful calibration steps for every deployment.