Professor Irene Lena Hudson, Department of Statistics, Data Science & Epidemiology, Swinburne University of Technology, Vic
I will discuss dynamic state space models that allow the estimation of the state of a phenological system given a multivariate time series of observations. These models facilitate stochastic trend detection allowing for time varying trends in contrast to studies that rely on time invariant trends, which are inclined to incorrectly estimate actual trends underlying phenological observations. For example contrasting periods of climate cooling and climate warming have manifested as temporal sign-switching in phenological trends, which are considered a strong fingerprint of a biotic response to climate change. The models I discuss within a unified framework accommodate for explicit modelling of the trends, non-stationary time series (in that the mean and variance may change over time), missing data and for the separation of the underlying process error (due to changing environmental conditions) from the observation error (due to differences in training or methods adopted by the phenological observers). The methodology allows for investigation simultaneously of non-linear and lagged relationships and the creation of pooled estimates of multi-parameter associations, that is meta-analytical estimates of interpretable summaries from complex non-linear and delayed associations.
This methodology has application to modelling for example; reproductive seasonality in relation to rainfall; acceleration of phenological advances with varying latitude/longitude/region; changes in timing of avian spring migration over time and space in relation to temperature; and effects of diurnal temperature range in mortality in urban settings. Likewise the models link with an approach reported by Hudson (2018) where the timing of a phenophase is influenced by a previous phenophase (state), e.g. rate of bud development may influence the timing and quantity of flowering in Eucalypts as does climate.