Modeling urban growth using video prediction technology: A time-dependent convolutional encoder-decoder architecture

摘要

This paper presents a novel methodology for urban growth prediction using a machine learning approach. The methodology treats successive historical satellite images of an urban area as a video for which future frames are predicted. It adopts a time-dependent convolutional encoder-decoder architecture. The methodology's input includes a satellite image for the base year and the prediction horizon. It constructs an image that predicts the growth of the urban area for any given target year within the specified horizon. A sensitivity analysis is performed to determine the best combination of parameters to achieve the highest prediction performance. As a case study, the methodology is applied to predict the urban growth pattern for the Dallas-Fort Worth area in Texas, with focus on two of its counties that observed significant growth over the past decade. The methodology is shown to produce results that are consistent with other growth prediction studies conducted for the areas.