Predicting Poverty Level From Satellite Imagery Using Ensemble Learning

Abstract

In order to effectively reduce poverty, it is essential to measure and follow support initiatives over time in order to focus aid efforts and inform policy choices. However, gathering such data requires a lot of time and effort, thus coverage of places plagued by poverty is frequently scant or nonexistent. Previous studies have demonstrated the viability of using remote sensing techniques to measure poverty levels. Particularly, convolutional neural network processing of satellite pictures has demonstrated potential in forecasting the intensity of nocturnal lights, which may then be used to determine the underlying poverty level. By figuring out ways to gauge changes in poverty levels over time using the same kind of readily accessible data, this initiative aims to build on earlier research. We are able to confirm the initial findings of the single-point poverty prediction. To meaningfully anticipate temporal poverty, further work is still required. In order to find interventions for projects to reduce poverty and equitably allocate resources, it is essential to ascertain the levels of poverty in different regions of the world. However, it is difficult to find accurate information on global economic conditions, particularly for regions in the developing world. This hinders efforts to both implement services and monitor/evaluate success. The goal of this research is to use satellite imagery to identify economic activity and, as a result, gauge the level of poverty in a certain area. A recurrent neural network is trained to understand several development characteristics, such as the type of rooftop, the illumination source, the distance from water sources, agricultural areas, and industrial areas.