A Bayesian maximum entropy reconstruction of stock distribution and inference of stock density from line-transect acoustic-survey data

Abstract

AbstractWe present a maximum-entropy (MaxEnt) method for inferring stock density and mapping stock distribution from acoustic line-transect data. MaxEnt is founded on the bedrock of probability theory and allows the most efficient possible use of known data in the inference process. The method takes explicit account of spatial correlation in the observed data and seeks to reconstruct a distribution of density across the whole survey area that is both consistent with the observed data and for which the entropy is maximized. The method is iterative and uses the Bayesian approach of evaluating the posterior probability of a candidate solution under the constraint of the observed data to progress towards a converged solution. We apply the method to reconstruct maps of the distribution of Antarctic krill throughout areas 100 × 80 km2. Survey data were integrated at 0.5-km intervals along ten 80-km transects, giving approximately 1600 observed data points. We inferred the krill density for all 32 000 0.5 × 0.5 km2 cells in the area. The method is demanding computationally, but appears to work well even in cases when the distribution of density is highly skewed. The MaxEnt technique has proved to be powerful for reconstructing quantitative images from incomplete and noisy physical data (e.g. radio-telescope data): we suggest that it has utility in biological systems too and could be of benefit to the fisheries-acoustics community, increasing the accuracy of acoustic estimates of stock density and generating better maps of stock distribution.