A0529
Title: Evaluating alternative deep learning approaches for village-level wealth estimation using satellite imagery
Authors: David Newhouse - World Bank Group (United States) [presenting]
Diana Jaganjac - World Bank Group (United States)
Josh Merfeld - University of Queensland (Australia)
Kushan Weerakoon - ProDex labs (United States)
Abstract: Building on existing research that uses satellite imagery and auxiliary data to estimate poverty at hyperlocal levels, transformer architectures and convolutional neural networks are evaluated to generate estimates of mean asset index values at the enumeration area level in Malawi. Estimates are generated using Planet Imagery and evaluated in held-out test sets after combining two household surveys: The Integrated Household Survey and the Multiple Indicator Cluster Survey. Estimates generated using Resnet 18 and the first version of the Prithvi foundational model outperform other architectures, achieving out of sample Pearson correlations of approximately 0.81. This exceeds performance from the recently developed ConvNeXt convolutional neural network and two standard vision transformer models. A robustness check using the Prithvi model with lower-resolution Landsat imagery achieves an out-of-sample correlation of 0.71. The results indicate that large-scale utilization of new foundational models to combine household survey data and satellite imagery offers a promising approach to generating accurate village-level estimates of wealth indices in this context.
