Parallelization of ensemble neural networks for spatial land-use modeling

Artificial neural networks have been widely applied to spatial modeling and knowledge discovery because of their high-level intelligence and flexibility. Their highly parallel and distributed structure makes them inherently suitable for parallel computing. As the technology of parallel and high-performance computing evolves and computing resources become more widely available, new opportunities exist for spatial neural network models to benefit from this advancement in terms of better handling computational and data intensity associated with spatial problems. In this study, we present a hybrid parallel ensemble neural network approach for modeling spatial land-use change. Our approach combines the shared-memory paradigm and the embarrassingly parallel method by leveraging the power of multicore computer clusters. The efficacy of this approach is demonstrated by the parallelization of Fuzzy ARTMAP neural network models, which have been extensively used in land-use modeling applications. We adopt an ensemble structure of neural networks to train multiple models in parallel and make use of the entire dataset simultaneously. We evaluate the proposed parallelization approach by examining performance variation of training datasets with alternative sizes. Experimental results reveal great potential of higher performance achievement when our hybrid parallel computing approach is applied to large spatial modeling problems.