Deep Learning Approach using Satellite Imagery Data for Poverty Analysis in Banten, Indonesia

Abstract

Satellite imageries data provides abundant geospatial features of infrastructures, land uses, land covers, and economic activity footprints that are potential for domainspecific tasks. In this study, we investigate the use of satellite imageries data as spatial-based proxy indicators in predicting the percentage of poverty in Banten Province, Indonesia using a deep learning approach. The poverty dataset is taken from the Village Potential Data Survey (PODES) 2018 results published by Statistics Indonesia (BPS) as the assumed ground-truth labels. Our finding reveals a correlation between the night-time light satellite imagery and the percentage of poverty, hence the regression model to predict the percentage of poverty is constructed using convolutional neural networks (CNN) architecture. The correlation between night-time image data and the percentage of poverty in each village is negative 52 percent under log transformation. Our proposed model generates a promising root mean squared error (RMSE) of 5.3023 which is potentially beneficial to support the construction and monitoring of poverty statistics in Indonesia.