ENSCO has been developing systems for imputing source locations using transport and dispersion modeling in combination with sample records of pollutant concentrations.

Goodness-of-agreement measures between modeled pollutant concentrations at fixed receptor locations and sample records at those locations provide the bases for likelihood calculations conditioned on hypothesized source location and release schedule.

Abstracting release schedules as Markov processes is reasonably justified from a physical/engineering perspective and permits exhaustive exploration of the enormous space of possible (unknown) release schedules using dynamic programming algorithms within a framework of hidden Markov models. The so-called Forward and Backward algorithms facilitate likelihood calculations marginalized over all possible state-paths (release schedules) of the hidden Markov process. Meanwhile, the Viterbi algorithm can be used to estimate the most-probable (a posterori) state-path.

We propose an estimation framework using hidden Markov models and demonstrate the approach using synthetic source/receptor data generated using transport and dispersion models ingesting historical wind fields over a several-month period in 2006.