Artificial neural networks and deep learning in urban geography: A systematic review and meta-analysis. (March 2019)

• Record Type: Journal Article
• Authors: Grekousis, George
• Is Part Of: Computers, environment and urban systems. Volume 74(2019)
• Page Start: 244
• Page End: 256
• Publication Date: 2019-03
• Subjects: Artificial Neural Networks -- Deep Learning -- Urban Geography -- Meta-analysis -- Trends -- Guidelines on Reporting results; City planning -- Data processing -- Periodicals; Regional planning -- Data processing -- Periodicals; 303.4834
• DOI: 10.1016/j.compenvurbsys.2018.10.008
• Languages: English
• ISSNs: 0198-9715

Abstract

Abstract: Artificial neural networks (ANNs) and their latest advancement in deep learning are blooming in computer science. Geography has integrated these artificial intelligence techniques, but not with the same enthusiasm. The main reason for hesitation is that ANNs are still confronted as complex and black boxes. However, ANNs might be more solid methods than conventional approaches when dealing with complex geographical problems. This study considers the great potential of ANNs for research in urban geography. First, using the PRISMA protocol, it provides a statistical review of 140 papers on studies that employed ANNs in urban geography between 1997 and 2016. Second, it performs a quantitative meta-analysis using non-parametric bootstrapping. 45 (of the 140) papers were assessed regarding ANNs' overall accuracy (OA) achieved when used for urban growth prediction or urban land-use classification. Third, a new guideline for reporting ANNs is proposed. Statistical review indicated that ANNs performed better in 75.7% of case studies compared to conventional methods. Meta-analysis found that on bootstrapped averages, the median OA achieved when using, ANNs was higher than the median OA achieved by other techniques by 2.3% ( p < .001). ANNs also performed better when used for classification compared to prediction. Analysis also identified inadequate presentation of ANNs and related results when used in urban studies. For this reason, a new guideline for reporting ANNs is suggested in this work to ensure consistency and easier dissemination of individual lessons learned. These findings aim to motivate further studies on ANNs and deep learning in urban geography. Graphical abstract: Highlights: ANNs performed better in 75.7% of case studies compared to conventional methods ANNs perform better when used for classification compared to prediction Major gap in integrating socio-economic data to ANNs is identified A guideline for systematic reporting of ANNs results and settings is suggested