Machine learning application to spatio-temporal modeling of urban growth. (June 2022)

• Record Type: Journal Article
• Authors: Kim, Yuna; Safikhani, Abolfazl; Tepe, Emre
• Is Part Of: Computers, environment and urban systems. Volume 94(2022)
• Publication Date: 2022-06
• Subjects: Urban growth -- Machine learning method -- Spatio-temporal modeling -- Random forest -- Prediction; City planning -- Data processing -- Periodicals; Regional planning -- Data processing -- Periodicals; 303.4834
• DOI: 10.1016/j.compenvurbsys.2022.101801
• Languages: English
• ISSNs: 0198-9715

Abstract

Abstract: Understanding the dynamics of urban growth is among the most important tasks in urban planning due to their influence on policy decision-making. Specifically, prediction of urban growth at regional levels is crucial for regional policy makers. Making such predictions is difficult because of the existence of complex topological structures and the high-dimensional nature of data sets related to urban growth. Spatial and temporal auto-correlation and cross-correlations, together with regional social and physical covariates, need to be properly accounted for improving the forecasting power of any statistical or machine learning method. To that end, we develop novel machine learning methodologies to perform predictions of urban growth at regional levels by incorporating lead-lag non-linear relationships among past urban changes in each region and its neighbors. Based on this analysis, machine learning algorithms outperform more classical methods, such as a logistic regression, in terms of classifying low/high urban growth regions, and the random forest algorithm seems to have the best prediction accuracy among the selected machine learning methods. Moreover, the random forest method without any external covariates has still a high prediction accuracy which not only confirms that most of variability of urban growth can be described by past observations of self and neighboring changes, but also makes it possible to perform real forecasting of urban growth without accessing any external covariates. The latter makes this modeling framework useful for local policy makers in allocating budget and directing resources appropriately based on such predictions. Highlights: Machine Learning applications are implemented to land-use change modeling Non-linear relationships in land developments are incorporated in models Both contemporaneous and historical dependencies are incorporated in machine learning models Random Forest model achieves almost 80% accuracy rate using only spatio-temporal lags of the dependent variable. The block group data is derived using approximately 9 million parcels between 1900 and 2019 for the state of Florida.