A Hybrid Framework for Space-Time Modeling of Environmental Data

The space-time autoregressive integrated moving average (STARIMA) model family provides useful tools for modeling space-time processes that exhibit stationarity (or near stationarity) in space and time. However, a more general method for routine use and efficient computation is needed to model the nonlinearities and nonstationarities of environmental space-time series. This article presents a hybrid framework combining machine learning and statistical methods to address this issue. It uses an artificial neural network (ANN) to extract global deterministic (nonlinear) space-time trends and a STARIMA model to extract local stochastic space-time variations in data. A four-stage procedure is proposed for analyzing and modeling space-time series. The proposed framework and procedures are applied to forecast annual average temperature at 137 national meteorological stations in China. The results demonstrate that the hybrid framework achieves better forecasting accuracy than the STARIMA model alone. This finding suggests that the combination of machine learning and statistical methods provides a very powerful tool for analyzing and modeling space-time series of environmental data that have strong spatial nonlinear and nonstationary components.