Mapping poverty with machine learning using publicly-available nightlights intensity, points-of-interest, and internet speed data
Access to high-resolution and up-to-date poverty maps can help organizations identify vulnerable populations and strategically prioritize and plan their poverty reduction interventions. Traditional data collection methods rely on resource-intensive and time-consuing household surveys, so available poverty data are limited in granularity and frequency, particularly for low- and middle-income countries like the Philippines. Recent work by Tingzon et al. [Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci. XLII-4/W19,
425 (2019)] has employed machine learning with publicly-available geospatial datasets to generate ~18 sq. km. resolution wealth index maps of the Philippines. Here, we expand on this work by including features from Internet speed data from Ookla, in addition to nightlights intensity features from NPP-VIIRS DNB satellite data and points-of-interest features from OpenStreetMap. We find that inclusion of the Internet speed features had a positive impact on model performance, translating to a slight improvement in the adjusted R2-score from 62.2% to 63.4%. Further work can utilize overlay these maps with population and demographic maps, electrification and Internet connectivity maps, crowd-sourced and citizen-generated maps, etc., to better understand geographical disparities in wealth and vulnerabilities across the country.