Abstract: In this paper, we use deep learning to estimate living conditions in India. We use both census and surveys to train the models. Our procedure achieves comparable results to those found in the literature, but for a wide range of outcomes.
Introduction
Traditional methods for measuring human development, such as national censuses and household surveys, are the "gold standards" of socio-economic data. However, they suffer from two major flaws: they are expensive and infrequent. In India, a census occurs only once a decade, leaving policymakers effectively "blind" to rapid local changes in the intervening years.
While previous research has utilized Earth Observation and Machine Learning (EO-ML) to fill these gaps, most studies have relied on nightlight luminosity as a proxy for wealth. This paper introduces a more granular approach, developing a deep-learning framework that uses daytime satellite imagery to estimate over 90 different health and material-living standards across rural India.
Methodology
The Asset-Transfer Innovation
The core innovation of this study is the move away from "nightlight" proxies toward a multi-dimensional "Asset-Transfer" model. Instead of training a model to predict light intensity, the researchers trained Convolutional Neural Networks (CNNs)—specifically ResNet and VGG architectures—to predict a 16-dimensional vector of material assets (such as housing quality and electricity access) derived from the Indian Census.
This approach allows the model to learn rich features from daytime imagery—textures, colors, and spatial patterns of human settlement—that are more indicative of development than simple luminosity. This feature-rich model is then used to predict "distal" outcomes: variables not directly visible from space, such as literacy rates or health markers, by leveraging the spatial correlation between the material environment and social outcomes.
Temporal Dynamics and Data Integration
The study integrated five massive data sources, including the 2001 and 2011 Indian Censuses, the National Family Health Survey (NFHS), and Landsat 7 imagery. To account for the fact that wealth indicators change over time (e.g., a mobile phone being a luxury in 2001 but ubiquitous in 2011), the researchers used advanced distribution-matching techniques like Linear Optimal Transport to correct for non-stationarity in the data.
Key Results
High Accuracy for Physical and Social Assets
The ResNet-34 model achieved significant predictive accuracy, particularly for physical infrastructure:
- Permanent Housing: $R^2$ of 0.69.
- Census Assets: Average $R^2$ of 0.5 across all 16 material indicators.
- Literacy: The asset-based model achieved an $R^2$ of 0.34, significantly outperforming the nightlight-based baseline of 0.21.
Predicting Health Outcomes
The model successfully bridged the gap to health metrics. For the 93 health outcomes analyzed, the model achieved an $R^2 \ge 0.5$ for 27 variables. Top-performing metrics included the percentage of women who are overweight and the frequency of births by caesarean section.
Discussion and Limitations
While the model provides a powerful new lens for development, it faces specific constraints. It struggled with "invisible" outcomes that lack a strong physical footprint or wealth correlation, such as vaccination rates or anemia prevalence. Furthermore, the study was geographically limited to six northern and eastern Indian states.
Future work will focus on expanding this framework to the entire Indian subcontinent and refining how assets are weighted to better reflect their contribution to human development. By providing a high-frequency, low-cost supplement to traditional surveys, this EO-ML approach offers a vital tool for real-time policy evaluation in developing nations.
Summary updated