Proceedings of the International Joint Conference on Neural Networks
An important feature present in neural network models is their ability to learn from data, even when the user does not have much information about the particular dataset. However, the most popular models do not perform well in spatial interpolation problems due to their difficulty in accurately modeling spatial correlation between samples. On the other hand, one of the most important geostatistical methods for spatial interpolation, Kriging, performs very well but requires some expert knowledge to fit the correlation model (semivariogram). In this work, we adapt the Incremental Gaussian Mixture Network (IGMN) neural network model for spatial interpolation and geostatistical sequential simulation applications. Results show that our approach outperforms Multilayer Perceptron (MLP) and the original IGMN, especially in anisotropic and sparse datasets. Also, we propose an algorithm for Sequential Gaussian Simulation that uses IGMN instead of Kriging and can successfully generate equally probable realizations of the defined grid. To the best of our knowledge, this is the first time a Neural Network model is specialized for spatial interpolation applications and has the ability to perform a geostatistical simulation. © 2016 IEEE.
Year of publication: 2016