The Ability of Artificial Neural Networks in Learning Dependency of Spatial Data

Abstract

In conventional methods of spatial data analysis, such as Kriging, the dependency structure of data is estimated, modeled, and then used for data prediction. In contrast, the Artificial Neural Network (ANN) approach, which is a data-driven approach, does not model the data dependency structure. Therefore, an important question may arise here: Does ANN use, indirectly, spatial dependency structure in data prediction? In this paper, we want to answer this question through a simulation study. Different dependent and independent spatial data sets are simulated under two spatial structures, and the prediction accuracy of ANNs is compared for simulated data. It is shown that neural network error for predicting dependent spatial data is much less than that of independent spatial data. We conclude that the network can indirectly learn spatial dependence between the observations. We also applied the ANN method to an experimentally obtained data set and compared its prediction accuracy with Kriging as a common geostatistical method. The results showed that the neural network can be used as an alternative method for spatial data prediction.