Reliable measurements of economic well-being at the local level are critical for effective public service delivery, policymaking, program evaluation, and private sector development. However, data on key economic indicators are scarce in many parts of the developing world, particularly in Africa. Nationally representative surveys, the main source of such data, are infrequent (at least 4 years between surveys in most African countries), have limited repeated observations of individual locations, and rarely release disaggregated data publicly. The infrequent nature of these surveys leads to a significant gap in understanding the dynamics of local-level economic well-being. This paper addresses this data scarcity by exploring the use of readily available, high-frequency satellite imagery as an alternative data source for measuring economic well-being. Past research has demonstrated the potential of nighttime lights imagery for measuring country-level economic performance and high-resolution imagery for assessing spatial variations in economic outcomes. This study aims to utilize more readily available, coarser-resolution multispectral satellite imagery, combining it with deep learning techniques to create spatially and temporally comprehensive estimates of economic well-being across sub-Saharan Africa, even in data-scarce regions.

Previous research has explored various methods for estimating economic well-being using remotely sensed data. Henderson et al. (2012) demonstrated the use of nighttime lights imagery to measure country-level economic performance over time. Studies like Jean et al. (2016) and Babenko et al. (2017) showed the potential of high-resolution imagery combined with machine learning for predicting poverty in specific regions. Other work has explored using mobile phone data (Blumenstock et al., 2015) and social media data (Sheehan et al., 2019) for similar purposes. However, these approaches often rely on either data that are not readily available or are limited in their spatial or temporal coverage. This study builds upon this prior research by employing publicly available, multispectral satellite imagery, which offers broader coverage and frequency, coupled with deep learning techniques, to improve the accuracy and scalability of the predictions.

This study employs a deep learning approach using publicly available multispectral satellite imagery and nighttime lights data to predict village-level asset wealth. The data used include asset wealth information for over 500,000 households from nationally representative Demographic and Health Surveys (DHS) conducted between 2009 and 2016 across 23 African countries. Asset wealth is calculated as a wealth index derived from the first principal component of survey responses on asset ownership. The satellite imagery consists of 30m/pixel Landsat multispectral imagery and <1 km/pixel nighttime lights imagery. These data are matched spatially and temporally with the survey data. A convolutional neural network (CNN) with a ResNet-18 architecture is used to predict the village-level wealth index. The model incorporates both daytime multispectral and nighttime lights imagery as inputs. The CNN model is trained using a combination of in-country and out-of-country training data to assess generalization performance. Model performance is evaluated using various metrics, including the R-squared value, which measures the proportion of variance in the ground-measured wealth explained by the model's predictions. For evaluating temporal variation, the study uses matched pairs of clusters from different survey years to assess wealth changes over time. Additionally, independent Living Standards Measurement Surveys (LSMS) data are used for validation. The study also explores a transfer learning approach where nighttime lights are used as intermediate labels for feature extraction from daytime imagery.

The combined model, utilizing both daytime multispectral and nighttime lights imagery, demonstrates strong predictive performance. It explains approximately 70% of the spatial variation in ground-measured village-level asset wealth in held-out countries, exceeding the performance of previous benchmarks using high-resolution imagery. The model's performance in individual countries is consistently high, ranging from above 50% to above 80% variation explained. Performance improves further when aggregating predictions to the district level (83% of variation explained). Analysis of the model's learned features indicates that it successfully identifies semantically meaningful features, such as urban areas, agricultural regions, water bodies, and deserts, which are intuitively related to wealth. Interestingly, models trained only on nighttime lights or multispectral imagery perform almost as well as the combined model, suggesting that both datasets contain similar information for spatial predictions. However, when assessing changes in wealth over time, models using multispectral daytime imagery outperform those using nighttime lights, likely due to the limited temporal variation in nighttime lights data. The study finds that the model can explain between 15% and 50% of the variation in survey-measured changes in asset wealth at the village and district levels respectively. A comparison with independent census data suggests that errors in satellite estimates are comparable to those in existing ground data. Furthermore, experiments show that the model's performance is affected by noise in the ground data, particularly the random displacement of GPS coordinates introduced for privacy reasons. The study demonstrates the utility of satellite-based estimates for research questions, such as examining the relationship between wealth and temperature exposure, and for policy tasks, such as targeting social protection programs. Finally, the scalability of the method is showcased by creating a wealth map for Nigeria.

This study successfully demonstrates the feasibility and accuracy of using publicly available satellite imagery and deep learning to estimate local-level economic well-being in Africa. The results show that the approach can generate accurate and scalable estimates, even in data-scarce regions, surpassing the accuracy of some previous methods. The ability of the model to generalize to countries not used in training is particularly encouraging, highlighting its potential for broad application across Africa. The findings also contribute to a growing body of research using remote sensing and machine learning to address challenges in development economics. While the model performs exceptionally well for spatial prediction, its interpretability could be enhanced for easier adoption by policymakers. Future research could focus on improving model interpretability while maintaining performance, addressing the limitations caused by noise in ground data, and expanding the application to other key economic indicators. The integration of higher-resolution imagery and other data sources, such as mobile phone or social media data, could further improve prediction accuracy.

This paper presents a novel approach to measuring economic well-being at the local level in Africa, using readily available satellite imagery and advanced deep learning techniques. The method outperforms previous benchmarks and proves scalable for use in large geographic areas. This offers a powerful tool for researchers and policymakers seeking to better understand and address poverty and inequality. Future work should focus on enhancing model interpretability and expanding its applications to other relevant development indicators. The approach has the potential to significantly improve the availability and timeliness of data for informed decision-making.

The study acknowledges limitations associated with the quality and availability of ground truth data. The random displacement of GPS coordinates in the DHS data introduces noise and reduces the accuracy of the model's predictions. The study’s findings on temporal changes in wealth are limited by the infrequent and non-panel nature of the available survey data. Furthermore, while the model is highly performant, it requires additional interpretation to be optimally used for policy interventions. Finally, the model's reliance on satellite imagery might exclude some aspects of economic well-being that are not captured in visual data.