Bayesian likelihood-free localisation of a biochemical source using multiple dispersion models

Branko Ristić, Ajith Gunatilaka, Ralph Gailis, Alex Skvortsov

Abstract

Localising the source of a toxic release of biochemical aerosols in the atmosphere is a critical challenge for public safety. Two primary difficulties hinder effective localisation: the lack of a known likelihood function for measurements collected by biochemical sensors, and the vast array of candidate dispersion models developed under varying meteorological and terrain assumptions. To address these challenges, this paper proposes a likelihood-free Approximate Bayesian Computation (ABC) method. This framework simultaneously utilizes a set of candidate dispersion models to localise the source, implemented via a Monte Carlo method and validated against two experimental datasets.

Introduction

When a sensor network detects a toxic aerosol in an urban or open environment, emergency responders must immediately identify the source location to mitigate the threat. This is fundamentally an "inverse problem": given the observed effects (sensor measurements), one must determine the cause (the source location and strength).

Traditional Bayesian approaches to this problem are often stymied by two major hurdles:

  1. Likelihood Uncertainty: Standard methods require a precise mathematical description—a likelihood function—of how sensor noise and modeling errors behave. In real-world scenarios involving turbulent winds and fluctuating concentrations, this function is often "intractable" or unknown.
  2. Model Uncertainty: No single physics model perfectly captures atmospheric dispersion across all conditions. Different models, such as Gaussian plumes or stretch exponential models, perform better in specific terrains or weather conditions, but the optimal model is rarely known a priori.

Methodology: Likelihood-Free ABC

To bypass the need for an explicit likelihood function, the authors employ Approximate Bayesian Computation (ABC). ABC shifts the focus from complex mathematical derivations to simulation. The core logic follows a simple iterative process: a candidate source location and dispersion model are selected, a simulation is run to predict sensor readings, and these predictions are compared to actual observed data. If the simulated data is "close enough" to the real data (within a tolerance threshold $\epsilon$), the candidate is retained.

The Adaptive Iterative Multiple-Model ABC Sampler

The paper introduces a sophisticated Adaptive Iterative Multiple-Model ABC Sampler that improves upon standard ABC in two ways:

  • Model Averaging: The system runs three dispersion models in parallel—two variations of the Gaussian Plume and one Stretch Exponential model. Rather than forcing a choice between them, the algorithm treats the model itself as a random variable, assigning probabilities based on how well each model matches the observed data.
  • Adaptive Iteration: Instead of a static tolerance threshold, the algorithm begins with a loose $\epsilon$ and gradually tightens it over successive iterations. This "anytime algorithm" approach ensures that even if the process is interrupted early, a preliminary estimate is available, which is vital for emergency response.

Experimental Validation

The framework was tested using experimental data from the COANDA Research & Development Corporation, which utilized a large water channel to simulate atmospheric dispersion at scale. Two distinct scenarios were analyzed:

  1. Open Terrain (Dataset 1): 48 sensors arranged in a grid with no obstacles, mimicking a flat field.
  2. Urban Scenario (Dataset 2): The same grid but with 10mm high obstacles placed to simulate the turbulence and complexity of a city environment.

Key Insights and Results

Model Selection and Complexity

One of the most significant findings was that model complexity does not always equate to accuracy. In the open terrain test, the probability shifted decisively toward a simpler 9-parameter Gaussian model ($m=2$), eventually dominating the more complex 10-parameter Stretch Exponential model. This suggests that the algorithm effectively identifies the most "credible" model for the specific environment based on the data provided.

Performance in Complex Environments

In the Open Terrain scenario, the algorithm was highly successful, with the posterior distribution clustering tightly around the true source location after 10 iterations.

However, the Urban Scenario highlighted the challenge of "model mismatch." The presence of obstacles created a bi-modal posterior, where the algorithm identified two likely source locations. While one mode was near the true source, the results were slightly biased. This occurred because none of the candidate models perfectly accounted for the specific turbulence patterns created by the urban obstacles, demonstrating that the accuracy of ABC is still bounded by the quality of the available simulators.

Computational Efficiency

The iterative approach proved highly efficient, maintaining an average sample acceptance rate of approximately 2.8%–2.9%. This is a significant improvement over standard non-iterative ABC methods, which often suffer from extremely low acceptance rates as the tolerance threshold tightens.

Conclusion

This research demonstrates that likelihood-free Bayesian methods provide a robust framework for source localisation when mathematical models are incomplete or sensor noise is unpredictable. By integrating multiple dispersion models and an adaptive sampling strategy, the system provides a flexible tool for emergency response. Future work will focus on incorporating more sophisticated models for urban turbulence and exploring the utility of the framework with simpler, binary "detect/no-detect" sensors.

Summary updated

BibTeX
@inproceedings{ristic2014bayesian,
  author = {Branko Ristić and Ajith Gunatilaka and Ralph Gailis and Alex Skvortsov},
  title = {{Bayesian likelihood-free localisation of a biochemical source using multiple dispersion models}},
  booktitle = {Signal},
  year = {2014},
  doi = {10.1016/j.sigpro.2014.08.023},
}
Figure