Granularity at Scale: Estimating Neighborhood Socioeconomic Indicators From High-Resolution Orthographic Imagery and Hybrid Learning

Abstract

Many areas of the world are without basic informa-1 tion on the socioeconomic well-being of the residing population 2 due to limitations in existing data collection methods. Overhead 3 images obtained remotely, such as from satellite or aircraft, 4 can help serve as windows into the state of life on the ground 5 and help "fill in the gaps" where community information is 6 sparse, with estimates at smaller geographic scales requiring 7 higher resolution sensors. Concurrent with improved sensor reso-8 lutions, recent advancements in machine learning and computer 9 vision have made it possible to quickly extract features from 10 and detect patterns in image data, in the process correlating 11 these features with other information. In this work, we explore 12 how well two approaches-a supervised convolutional neural 13 network and semi-supervised clustering based on bag-of-visual-14 words-estimate population density, median household income, 15 and educational attainment of individual neighborhoods from 16 publicly available high-resolution imagery of cities throughout 17 the United States. Results and analyses indicate that features 18 extracted from the imagery can accurately estimate the density 19 (R 2 up to 0.81) of neighborhoods, with the supervised approach 20 able to explain about half the variation in a population's income 21 and education. In addition to the presented approaches serving 22 as a basis for further geographic generalization, the novel semi-23 supervised approach provides a foundation for future work 24 seeking to estimate fine-scale information from aerial imagery 25 without the need for label data.