Deep Learning with Satellite Images Enables High-Resolution Income Estimation: a Case Study of Buenos Aires

This repository contains the datasets required for replicating the results in Abbate et al (forthcoming). The datasets also include per capita income estimates at a 50x50 meter resolution for the years 2013, 2018, and 2022, using satellite images from the Metropolitan Area of Buenos Aires (Argentina) and 2010 census+survey data. The model, based on the EfficientnetV2 architecture, achieved high accuracy in predicting household incomes (Rsquared=0.878), surpassing existing methods in spatial resolution and performance. 

Inside the Replication Package folder, the user can replicate the main results from the paper. This includes:

1. Small Area Estimation (SAE) Replication:

   • Argentina Household Survey Data (EPH): Processed microdata for 2010, 2013, 2018, and 2022 (ARG_*_EPHC-S2_*.dta).

   • Argentina Census Microdata: Raw 2010 census microdata (censo2010_fullraw_p.dta).

   • Census Tract Map: Shapefile of 2010 census tracts (radios_eph_with_link.shp).

   • SAE Output: The final small_area_estimates.parquet file containing census tract-level population and estimated income, which serves as labels for the CNN model.

2. CNN-based Income Prediction Replication (Paper Results):

   • CNN Model Income Predictions: Gridded 50x50m income estimates for Buenos Aires for 2013, 2018, and 2022 (income_estimates_*.shp).

   • Normalization Scalars: A CSV file (scalars_ln_pred_inc_mean_trimTrue.csv) to convert the model's log-scale outputs into real income values (2010 PPP-adjusted Argentinian pesos).

   • World Settlement Footprint (WSF): Satellite-based data (WSF2015_v2_-60_-36.tif) used to mask predictions in uninhabited areas.

Key prediction datasets are published in shapefile format, while input data for SAE and other auxiliary files are in formats like .dta, .parquet, .csv, and .tif.

Results can be replicated by connecting these datasets with the scripts available at the GitHub repo linked below.

Important Usage Note: Since the predictions for each 50x50m cell individually present some random variation, we recommend that the results are used by averaging out the estimations for each area of interest (e.g., municipalities, neighborhoods, sections, or census tracts) and not at an individual cell level. As detailed throughout the paper, the aggregated results, even in small areas such as census tracts, predict household incomes with precision.

Furthermore, inside this repository, it is possible to access and use the model’s trained parameters to make predictions about different satellite images.

Data can be visualized by accessing: https://ingresoamba.netlify.app