[Paper Review] On monitoring development indicators using high resolution satellite images

P. Suraj, Ankesh Gupta+3 more|arXiv (Cornell University)|Dec 6, 2017

Impact of Light on Environment and Health34 references4 citations

TL;DR

This paper proposes a deep learning framework using high-resolution daytime satellite images to predict socio-economic indicators at the village level in India, outperforming nightlight-based proxies. The method employs a custom 8-layer CNN (VGG-CNN-S) trained on 218,000 villages across six states, achieving superior R² scores (0.65 on average) and enabling accurate prediction, data validation, transfer learning for diverse indicators, and anomaly detection in development trends.

ABSTRACT

We develop a machine learning based tool for accurate prediction of socio-economic indicators from daytime satellite imagery. The diverse set of indicators are often not intuitively related to observable features in satellite images, and are not even always well correlated with each other. Our predictive tool is more accurate than using night light as a proxy, and can be used to predict missing data, smooth out noise in surveys, monitor development progress of a region, and flag potential anomalies. Finally, we use predicted variables to do robustness analysis of a regression study of high rate of stunting in India.

Motivation & Objective

• To develop a machine learning tool that accurately predicts socio-economic indicators from high-resolution daytime satellite imagery, especially where survey data are infrequent or unreliable.
• To overcome limitations of traditional proxies like nightlight data, which are less accurate and not always correlated with on-the-ground development metrics.
• To create a scalable, generalizable model for predicting diverse indicators—assets, infrastructure, demographics, education, and health—across geographically and culturally diverse regions in India.
• To demonstrate the utility of predicted variables in improving econometric studies, particularly in reducing omitted variable bias in regression analysis.
• To enable monitoring of regional development progress and detection of spatial anomalies through regression output discontinuities.

Proposed method

• A custom 8-layer deep convolutional neural network (VGG-CNN-S) is trained to directly regress village-level socio-economic indicators from high-resolution daytime satellite images.
• The model is trained on a dataset of 218,000 villages across six Indian states—Punjab, Haryana, Uttar Pradesh, Bihar, Jharkhand, and West Bengal—excluding urban areas.
• The framework uses transfer learning to predict additional socio-economic indicators (e.g., literacy, health, education) from the learned asset representation, improving generalization.
• Out-of-sample R² scores are used to evaluate performance, with the model achieving an average R² of 0.65 across multiple indicators.
• The model enables village-level prediction even from cross-sectional data, allowing temporal monitoring of development trends.
• Spatial discontinuities in regression outputs are analyzed as potential alerts for policy-relevant disparities between neighboring regions.

Experimental results

Research questions

• RQ1 Can high-resolution daytime satellite imagery be used to predict a wide range of socio-economic indicators more accurately than existing proxies like nightlight data?
• RQ2 To what extent can a single deep learning model trained on asset indicators be transferred to predict other socio-economic and health-related variables?
• RQ3 How can predicted variables from satellite imagery improve the robustness of econometric studies, particularly in reducing omitted variable bias?
• RQ4 Can spatial discontinuities in model predictions reveal meaningful regional disparities or policy anomalies?
• RQ5 Can the model be used to validate and smooth noisy census data, and to monitor development progress over time?

Key findings

• The proposed CNN-based model achieves an average out-of-sample R² score of 0.65 for predicting socio-economic indicators, significantly outperforming nightlight-based proxies.
• The model effectively smooths out noise in census data and serves as a validation tool, improving data reliability across diverse geographic and cultural regions.
• Transfer learning from the asset prediction model enables accurate prediction of literacy, education, health, and demographic indicators with high generalization capability.
• Village-level regression using predicted data reveals that women’s education is the most significant determinant of child stunting, with open defecation and infrastructure factors also playing key roles.
• Spatial discontinuities in regression outputs can signal policy-relevant disparities between neighboring regions, serving as early alerts for further investigation.
• The use of predicted variables in regression analysis reduces omitted variable bias and strengthens causal inference, as demonstrated in a robust case study on stunting determinants.