A Dirichlet process model for change-point detection with multivariate bioclimatic data

Motivated by real-world data of monthly values of precipitation, minimum, and maximum temperature recorded at 360 monitoring stations covering the Italian territory for 60 years (12x60 months), in this work we propose a change-point model for multiple multivariate time series, inspired by the hierarchical Dirichlet process. We assume that each station has its change-point structure and, as main novelties, we allow unknown subsets of the parameters in the data likelihood to stay unchanged before and after a change-point, that stations possibly share values of the same parameters and that the unknown number of weather regimes is estimated as a random quantity. Owing to the richness of the formalization, our proposal enables us to identify clusters of spatial units for each parameter, evaluate which parameters are more likely to change simultaneously, and distinguish between abrupt changes and smooth ones. The proposed model provides useful benchmarks to focus monitoring programs regarding ecosystem responses. Results are shown for the whole data, and a detailed description is given for three monitoring stations. Evidence of local behaviors includes highlighting differences in the potential vulnerability to climate change of the Mediterranean ecosystems from the Temperate ones and locating change trends distinguishing between continental plains and mountain ranges.