Spatio-temporal modeling of parcel-level land-use changes using machine learning methods

Abstract

Spatio-temporal modeling of parcel-level land development dynamics is essential to maintain sustainable urban growth. Modeling parcel-level urban development controlling contemporaneous and historical conditions involve computational challenges since data sizes grow quickly beyond the capabilities of conventional statistical-based spatial models. Machine Learning (ML) methods provide computationally feasible methods for large-scale data sets. This paper introduces new ML applications using advanced algorithms and GPU parallel processing to model large-scale urban land developments. Special attention is given to accelerating the construction of spatial weight matrices and training ML models. Specifically, artificial neural networks and random forests are applied to the state of Florida’s land-use data, which contains nearly 9 million parcels, to predict parcels with changes in their land use based on historical and neighborhood data. The adaptive Hashing algorithm coupled with GPU parallel processing accelerates the average processing time for identifying the fixed number of nearest neighbors used for accounting spatial autocorrelation, by almost 16,000 times. Also, ML model training times are shortened by 49–547 times using GPU. Further, our best ML model achieves approximately 92% accuracy while outperforming some competing methods, including logistic regression. Such a high prediction accuracy helps policymakers adjust budget allocations to meet local land-use change projections.