Predictive Performance of Kriging and the Long Short-Term Memory for Spatio-Temporal Data

Spatio-temporal data arise in many studies across agriculture, environmental, and climate sciences, where forecasting and prediction are often key objectives. The best linear unbiased prediction method, Kriging, is widely used in these fields. However, the large covariance matrix involved in Kriging poses significant computational challenges. Moreover, Kriging of-ten depends on model assumptions such as stationarity. Machine learning methods offer an alternative approach. In this work, we compare the predictive performance of a particular deep learning model-the Long Short-Term Memory (LSTM) network-with Kriging. Using two environmental datasets, we demonstrate that the LSTM has the potential to outperform Kriging, likely due to the fewer assumptions required by the LSTM model