Search

Several methods have been proposed for using satellite imagery to model poverty. These include poverty mapping using convolutional neural networks applied either directly or using transfer learning to high resolution satellite images, or combinations of methods that combine satellite imagery with standard methods. However, these methods require proprietary imagery which, given their cost and infrequent acquisition, may render these advances impractical for most applications. The authors investigate how satellite-derived poverty maps may improve when incorporating features derived from Sentinel-2 and MODIS imagery, which are both open-source and freely and readily available. The authors estimate a poverty map for Belize which incorporates spatial and time series features derived from these sensors, with and without survey derived variables. They document an 8% percent improvement in model performance when including these satellite features and conclude by arguing that Open Data for Development should include open data pipelines where possible. [ABSTRACT FROM AUTHOR]

With an increasing global population, accurate and timely population counts are essential for urban planning and disaster management. Previous research using contextual features, using mainly very-high-spatial-resolution imagery (<2 m spatial resolution) at subnational to city scales, has found strong correlations with population and poverty. Contextual features can be defined as the statistical quantification of edge patterns, pixel groups, gaps, textures, and the raw spectral signatures calculated over groups of pixels or neighborhoods. While they correlated with population and poverty, which components of the human-modified landscape were captured by the contextual features have not been investigated. Additionally, previous research has focused on more costly, less frequently acquired very-high-spatial-resolution imagery. Therefore, contextual features from both very-high-spatial-resolution imagery and lower-spatial-resolution Sentinel-2 (10 m pixels) imagery in Sri Lanka, Belize, and Accra, Ghana were calculated, and those outputs were correlated with OpenStreetMap building and road metrics. These relationships were compared to determine what components of the human-modified landscape the features capture, and how spatial resolution and location impact the predictive power of these relationships. The results suggest that contextual features can map urban attributes well, with out-of-sample R2 values up to 93%. Moreover, the degradation of spatial resolution did not significantly reduce the results, and for some urban attributes, the results actually improved. Based on these results, the ability of the lower resolution Sentinel-2 data to predict the population density of the smallest census units available was then assessed. The findings indicate that Sentinel-2 contextual features explained up to 84% of the out-of-sample variation for population density. [ABSTRACT FROM AUTHOR]