Distributed minimum temperature prediction using mixtures of gaussian processes

Abstract

Minimum temperature predictions are required for agricultural producers in order to assess the magnitude of potential frost events. Several regression models can be used for the estimation problem at a single location but one common problem is the amount of required data for training, testing and validation. Nowadays, sensor networks can be used to gather environmental data from multiple locations. In order to alleviate the amount of data needed to model a single site, we can combine information from the different sources and then estimate the performance of the estimator using hold-out test sites. A mixture of Gaussian Processes (MGP) model is proposed for the distributed estimation problem and an efficient Hybrid Monte Carlo approach is also proposed for the estimation of the model parameters.