Abstract:
|
The primary challenge of spatial statistics is to account for the inherent correlation that arises in the data due to proximity in sampling locations. To date, the go-to tool for spatial statisticians has been the Gaussian process (GP). However, the GP is plagued by computational challenges rendering it infeasible for use on modern, large spatial data sets. On the other hand, deep learning via deep neural networks has arisen as a flexible approach for modeling nonlinear relationships. To date, however, deep neural networks have only been scarcely used for problems in spatial statistics. In this work, we demonstrate how to implement deep neural networks for spatial data. We compare the deep neural networks with modern applications of Gaussian processes for various spatial data sets. In this comparison, we highlight the advantages and disadvantages of each approach.
|