[Paper Review] Poverty Prediction with Public Landsat 7 Satellite Imagery and Machine Learning

Anthony Perez, Christopher Yeh+4 more|arXiv (Cornell University)|Nov 10, 2017

Remote-Sensing Image ClassificationEngineering7 references51 citations

TL;DR

The paper uses publicly available Landsat 7 multi-spectral imagery and CNNs to predict poverty indicators in Africa, achieving competitive r^2 scores and better performance than prior Google-based imagery methods especially within poorer populations. It also analyzes cross-country generalization and temporal potential.

ABSTRACT

Obtaining detailed and reliable data about local economic livelihoods in developing countries is expensive, and data are consequently scarce. Previous work has shown that it is possible to measure local-level economic livelihoods using high-resolution satellite imagery. However, such imagery is relatively expensive to acquire, often not updated frequently, and is mainly available for recent years. We train CNN models on free and publicly available multispectral daytime satellite images of the African continent from the Landsat 7 satellite, which has collected imagery with global coverage for almost two decades. We show that despite these images' lower resolution, we can achieve accuracies that exceed previous benchmarks.

Motivation & Objective

• Motivate affordable, scalable poverty estimation in data-scarce developing regions using publicly available satellite data.
• Leverage Landsat 7 multi-spectral imagery to predict wealth indicators where surveys are sparse.
• Develop CNN architectures that handle multi-band imagery and varying spatial resolutions.
• Evaluate transfer learning from nighttime lights and derive poverty metrics from image features.

Proposed method

• Train CNNs (ResNet-18, ResNet-34, VGG-F) on Landsat 7 daytime multi-band images (9 bands) to predict nighttime light intensity classes.
• Adapt first convolutional layer to accept 9-band input and initialize RGB channels from ImageNet while averaging non-RGB bands across filters.
• Upsample to 15m resolution with dilated convolutions to align multi-band imagery and preserve pretrained weight usefulness.
• Extract image features from the CNNs and train regression or tree-based models (ridge, GBT) to predict the AWI poverty metric using DHS survey data.
• Apply leave-one-country-out cross-validation to assess cross-country generalization of poverty predictions.

Experimental results

Research questions

• RQ1 How well can public Landsat 7 multi-band imagery predict poverty indices (AWI) compared to night-time lights baselines?
• RQ2 Do multi-band Landsat features improve poverty prediction relative to RGB or nightlight-only models?
• RQ3 What is the cross-country generalization performance when predicting poverty for unseen countries?
• RQ4 Can temporal extension using Landsat 7 data enable tracking poverty changes over time?
• RQ5 How do different CNN architectures and training strategies affect predictive accuracy for poverty metrics?

Key findings

• CNNs trained on 9-band Landsat 7 imagery achieve higher out-of-country prediction r^2 than nightlights alone in several configurations.
• In leave-one-country-out settings, VGG-F with 9-band inputs and ridge regression achieves substantial predictive power, e.g., train r^2 0.68 and test r^2 0.63.
• ResNet-18 and ResNet-34 with 9-band inputs also yield strong predictions, e.g., train r^2 0.69–0.70 and test r^2 0.64–0.65.
• Compared to the previous Google Static Maps-based work, Landsat 7 9-band models outperform in pooled cross-validation, with aggregate residual r^2 around 0.70–0.74 for several models.
• Nightlights + GBT baseline achieves mean train r^2 0.63, mean test r^2 0.66, and aggregate residual r^2 1.0, indicating competitive performance but varying by model and setup.
• The results suggest stronger AWI predictions for poorer segments, and limited cross-country extrapolation, highlighting potential for within-country or pooled analyses over time.