Mapping Childhood Poverty: KidSat Project's Use of Satellite Imagery and Deep Learning

The KidSat project utilizes high-resolution satellite imagery combined with detailed survey data to map childhood poverty across Eastern and Southern Africa, offering a novel benchmark for evaluating deep learning models. This initiative aims to enhance poverty prediction accuracy, providing a scalable tool for policymakers to address child poverty effectively.

The KidSat project, an ambitious and collaborative effort by researchers from esteemed institutions such as the University of Oxford, University College London, the National University of Singapore, and Imperial College London, seeks to harness the power of satellite imagery to map childhood poverty across Eastern and Southern Africa. This initiative has resulted in the development of a groundbreaking dataset that combines satellite images with detailed survey data from 33,608 locations across 19 countries, covering a comprehensive period from 1997 to 2022. The primary objective of this dataset is to establish a benchmark for satellite feature representations, particularly in predicting child poverty, which is a crucial yet complex demographic indicator.

A New Benchmark for Child Poverty Mapping

Satellite imagery has increasingly become a critical resource for analyzing various demographic, health, and development indicators. Despite the proliferation of deep learning models designed for these tasks, the field has lacked standardized benchmarks, which are essential for evaluating and comparing the performance of these models. The KidSat dataset addresses this significant gap by pairing high-resolution satellite images with high-quality survey data on child poverty, as defined by UNICEF. Child poverty is a multifaceted issue measured across six dimensions: housing, water, sanitation, nutrition, health, and education. The dataset allows researchers to test models for both spatial and temporal generalization by evaluating their performance on unseen locations and data collected after the training years.

Leveraging Advanced Deep Learning Models

To benchmark the effectiveness of various models, the KidSat project utilized a range of models, from basic satellite imagery models like MOSAIKS to advanced deep learning foundation models such as DINOv2 and SatMAE. These models were tested on their ability to predict the percentage of children experiencing severe deprivation within a 10 km × 10 km area. The results indicated that fine-tuning transformer-based vision models significantly enhanced their performance in predicting child poverty. Notably, the DINOv2 model achieved the lowest mean absolute error (MAE) of 0.1836 when using Sentinel-2 imagery, underscoring the model's accuracy and reliability.

Understanding the Complexity of Child Poverty

The research underscores the proven effectiveness of remote sensing for tasks that are naturally visible from space, such as land usage prediction, crop yield forecasting, and deforestation. However, the KidSat project focuses on a more challenging task: multidimensional child poverty. This complexity arises because child poverty, unlike adult poverty, cannot simply be assessed by measuring overall household resources. Instead, it must be measured at the level of the child and their specific experiences, which include nutrition, health, and education needs. Failure to meet these needs can result in lifelong negative consequences. The internationally agreed definition of child poverty, designed to enable cross-country comparisons, includes crucial dimensions for children that require material resources, such as education, health, and nutrition.

Harnessing Satellite Imagery and Survey Data

The dataset incorporates high-resolution satellite imagery from the Landsat and Sentinel programs, chosen for their public accessibility and long history. For each specified survey coordinate, a 10 km × 10 km window of imagery was extracted using Google Earth Engine. The selection criteria for the imagery included the designation of a specific year and prioritization based on the least cloud cover within that year, ensuring the highest quality and suitability for accurate analysis. The Demographic and Health Surveys (DHS) Program, dating back to 1984, provided the survey data, offering up-to-date information on a wide range of demographic, health, and nutrition monitoring indicators. These surveys, funded by the US Agency for International Development (USAID), are conducted in partnership with country governments and provide nationally representative cross-sectional household data with very high response rates.

Implications for Policy and Future Research

In the spatial benchmarking, foundational vision models outperformed the baseline mean prediction model and Gaussian Process regression. Models like MOSAIKS, DINOv2, and SatMAE, particularly when fine-tuned with DHS variables, showed a reduction in mean absolute error. This indicates that spatial features extracted from satellite imagery are more effective than Gaussian Process modeling in estimating poverty indicators in regions where surveys have not been conducted. However, the temporal benchmark, which evaluated a forecasting task using data from before 2020 to predict poverty in 2020-2022, proved more challenging. Satellite imagery is at best a proxy for multidimensional child poverty, and this finding suggests it is a better proxy for quantifying spatial rather than temporal variation. Models performed worse on the temporal benchmark, and fine-tuned models, particularly those using Sentinel-2 imagery, showed increased MAE compared to the raw output. This suggests that models may overfit historical data, struggling to generalize to data collected after 2020.

The KidSat project highlights the potential of combining satellite imagery with deep learning to provide scalable, cost-effective methods for poverty estimation. This approach offers policymakers a valuable tool to make informed decisions and address child poverty in under-surveyed regions. The open-source code provided by the researchers allows for further exploration and application of these methods in other contexts, potentially driving significant advancements in socio-economic research and policy-making.