Combining environmental information from multiple and disparate data sources for minimum temperature prediction

Abstract

Minimum temperature forecasts are a highly valuable tool in agricultural decision making. In this paper, a method for statistical forecasting of minimum temperature predictions using multiple data sources is given. The proposed approach uses a well known method for non-linear regression such as Gaussian Process (GP) and produces a prediction for the minimum temperature for the next day. In addition to the forecast at a single spatial location, a mixture of GP (MGP) models is proposed in order to combine data from multiple independent data sources. The method is tested with real data collected from wireless sensor networks and preliminary results shows improved performance when using the MGP approach while also allowing a fast E-M algorithm for parameter estimation using the product of several predictive distributions.