Dynamic linear Bayesian models in phytoplankton ecology

Abstract

As phytoplankton time series show high variabilities which are generated by processes occurring at differing spatial and temporal scales, static regression models may not be adapted to the inherent complexity of the data. The aim of this paper was to consider the advantages of Bayesian dynamic models for phytoplankton time series. Dynamic models allow for time-varying influence of the covariates. The basic assumption is the existence of underlying and unobservable time series for the vector parameter whose distribution is sequentially estimated, allowing on-line analysis. The dynamic linear regression model (DLRM) is described and applied to a time series of the concentration of the marine toxic microalga Dinophysis cf. acuminata. The evolution in time of the regression parameters shows scales of influence in the environmental factors and provides a segmentation of the time series into significant and non-significant phases. In our application, physical factors accounted for most of the fluctuations in Dinophysis cf. acuminata concentrations. In particular, the wind parameter exhibited variations which could be interpreted as accumulation and dispersion phenomena. This kind of information may help in understanding the processes underlying the fluctuations in Dinophysis cf. acuminata concentrations, as long as a sensible interpretation can be found for the parameters evolution.