Exploring Spatial-Temporal Deep Learning Algorithms in Urban Expansion Prediction and Scenario Planning

Authors: Runzi Wang*, , Xiao Li, Texas A&M University, Haotian Zhong, Texas A&M University
Topics: Land Use and Land Cover Change, Quantitative Methods, Urban and Regional Planning
Keywords: Deep learning, machine learning, land-cover change, remote sensing, scenario planning
Session Type: Paper
Day: 4/9/2020
Start / End Time: 11:10 AM / 12:25 PM
Room: Tower Court A, Sheraton, IM Pei Tower, Second Floor Level
Presentation File: No File Uploaded


Urban land-cover change exaggerates the threats of climate change through loss of natural habitat and increases in carbon emission. However, little is known about future patterns and rates of urban expansion. Deep learning (DL) algorithms have emerged as a promising tool for urban land-cover change predictions. Combining the advantages of Convolutional Neural Networks and Recurrent Neural Networks, a convolutional LSTM (ConvLSTM) was proposed for capturing spatiotemporal correlation (Xingjian et al., 2015). Given the path-dependent nature of land-cover change, ConvLSTM presents an opportunity to improve the performance of land-cover change prediction.
In this study, we aim to 1) compare the performance of three spatial-temporal models in urban expansion prediction and 2) optimize the performance of ConvLSTM by leveraging ancillary data (including socioeconomic data, institutional contexts, and cultural structures) and geospatial constraints. We choose Mckinney, TX as our study site, which is among the nation’s fastest-growing cities from 2000 to 2010. A series of urban maps for the year 2001, 2003, 2006, 2008, 2011 will be used to train the DL models and test their performance in predicting an urban map in 2016. All the urban maps will be reclassified from NLCD. Candidate models for the comparison include a traditional CA-Markov model, fully connected LSTM (FC-LSTM), and ConvLSTM. We expect that by unlocking the potential of ancillary data and innovating on deep learning algorithms, the proposed research will contribute to the study of urban land-cover change by not only predicting future urban development but also constructing potential scenarios under different policy interventions.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login