High-resolution rural poverty mapping in Pakistan with ensemble deep learning

High resolution poverty mapping supports evidence-based policy and research, yet about half of all countries lack the survey data needed to generate useful poverty maps. To overcome this challenge, new non-traditional data sources and deep learning techniques are increasingly used to create small-area estimates of poverty in low- and middle-income countries (LMICs). Convolutional Neural Networks (CNN) trained on satellite imagery are emerging as one of the most popular and effective approaches. However, the spatial resolution of poverty estimates has remained relatively coarse, particularly in rural areas. To address this problem, we use a transfer learning approach to train three CNN models and use them in an ensemble to predict chronic poverty at 1 km(2) scale in rural Sindh, Pakistan. The models are trained with spatially noisy georeferenced household survey containing poverty scores for 1.67 million anonymized households in Sindh Province and publicly available inputs, including daytime and nighttime satellite imagery and accessibility data. Results from both hold-out and k-fold validation exercises show that the ensemble provides the most reliable spatial predictions in both arid and non-arid regions, outperforming previous studies in key accuracy metrics. A third validation exercise, which involved ground-truthing of predictions from the ensemble model with original survey data of 7000 households further confirms the relative accuracy of the ensemble model predictions. This inexpensive and scalable approach could be used to improve poverty targeting in Pakistan and other low- and middle-income countries.