Authors: Minrui Zheng*, University of North Carolina - Charlotte, Wenwu Tang, University of North Carolina at Charlotte, Xiang Zhao, Wuhan University
Topics: Geographic Information Science and Systems, Spatial Analysis & Modeling, Cyberinfrastructure
Keywords: neural network, machine learning, hyperaprameter optimization, spatial statistics
Session Type: Paper
Start / End Time: 1:10 PM / 2:50 PM
Room: Roosevelt 4, Marriott, Exhibition Level
Presentation File: No File Uploaded
Artificial neural networks (ANNs) as a machine learning approach have been extensively used for the spatially explicit modeling of complex geographic phenomena. However, because of complexity of computational process, there has been inadequate investigation on configuration of architectures of neural networks, such as the number of hidden layers, the number of nodes in each hidden layer, or learning rate. Given a specific spatial dataset, the performance of a model driven by ANNs often depends on the parameter setting of the model. But, most studies in the literature rely on a manual trial-and-error approach to select the parameter setting for ANN-driven spatial models. In this study, we develop an automated selection approach to identify optimal neural networks for spatial modeling using hyperparameter optimization. Hyperparameter optimization provides support for selecting the optimal architectures of ANNs. Yet, the use of hyperparameter optimization is often challenging because hyperparameter space is often large and the associated computational demand is heavy. Thus, we utilize high-performance and parallel computing to accelerate the model selection process. The spatial model used in our case study is a land price evaluation model that estimates the impact of driving factors on land price in an urbanized county: Mecklenburg County, North Carolina. Our results demonstrate that the automated selection approach improves the model-level performance compared with linear regression, and high-performance and parallel computing is of great help for accelerating the selection of optimal neural networks for spatial modeling.