Deep Learning for Poverty Estimation in Côte d’Ivoire using Sentinel-2 Imagery

K. M. Koné${}^a$, M. R. Kouamé${}^a$, F. A. Migoné${}^a$, D. J. Koné${}^a$, K. A. Kouassi${}^a$ and D. Doukouré${}^a$

${}^a$National Statistics Agency (ANStat), Côte d’Ivoire

This study develops a methodology for estimating poverty in Côte d’Ivoire using Sentinel-2 satellite imagery and deep learning transfer techniques. The approach follows a three-stage pipeline: pre-training a VGG16 convolutional neural network [1] on ImageNet [2], fine-tuning it to predict nighttime light intensity from daytime imagery, and extracting visual features for poverty prediction. Using data from the 2018-2019 Harmonized Household Living Conditions Survey covering 12,992 households, visual features are enriched with spectral indices including NO${}_2$ concentration, Land Surface Temperature, and the Enhanced Normalized Difference Impervious Surface Index. The best model achieves an $R^2$ of 40.15%, with predictions remarkably consistent with survey data: 39.98% predicted versus 39.44% observed in 2018. Applying the 2018-trained model to 2021 imagery yields an inferred rate of 37.86%, closely matching the 37.50% survey estimate, confirming the approach’s capacity to capture poverty dynamics over time. This methodology offers a cost-effective complementary tool for monitoring poverty between survey waves in Sub-Saharan Africa.

Keywords: Poverty estimation, Satellite imagery, Deep learning.