Authors: Steven Chao*, The George Washington University, Ryan Engstrom, The George Washington University, Michael Mann, The George Washington University, Adane Bedada, The George Washington University
Topics: Spatial Analysis & Modeling, Remote Sensing, Urban Geography
Keywords: machine learning, contextual features, population, urban attributes, modeling, spatial resolution
Session Type: Paper
Start / End Time: 9:35 AM / 10:50 AM
Room: Virtual Track 8
Presentation File: No File Uploaded
With an increasing global population, accurate and timely population counts are essential for urban planning and disaster management. Researchers have modeled population and socio-economic variables with contextual features derived from satellite imagery. We define contextual features as the statistical quantification of edge patterns, pixel groups, gaps, texture, and the raw spectral signatures calculated over groups of pixels or neighborhoods. Previous research using contextual features has mainly used very-high spatial resolution imagery at subnational to city scales and has found strong correlations with population and poverty. This study evaluates the feasibility and accuracy of using contextual features derived from multi-scale satellite imagery to model elements of the human-modified landscape. We calculate contextual features from very-high spatial resolution (<2m pixels) imagery and lower spatial resolution Sentinel-2 (10m pixels) imagery in Sri Lanka and Belize and correlate those outputs with OpenStreetMap road and building values. We then compare these relationships to determine how spatial resolution impacts the predictive power and how different countries affect the relationship. Finally, we assess the ability to predict the population density of the smallest census units available with Sentinel-2 contextual features. Our results suggest that the contextual features are able to map urban attributes well, with r-squared values ranging from 41% to 81%. Moreover, the degradation of spatial resolution does not significantly reduce the results, and for some variables, the results actually improved. The findings also indicate that Sentinel-2 contextual features can explain up to 74% of the variation in population density in our study area.