Human bias and CNNs' superior insights in satellite based poverty mapping

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Socioeconomic Studies

Background:

• Satellite imagery offers a valuable alternative for poverty estimation in data-scarce regions.
• Traditional methods often rely on expert interpretation, which can be subjective and limited.
• Machine learning approaches present opportunities for more objective and scalable poverty assessment.

Purpose of the Study:

• To compare expert-based and machine learning (ML) methodologies for poverty estimation using satellite imagery.
• To evaluate the performance of a convolutional neural network (CNN) against expert-defined features.
• To assess the utility of ML explainability in identifying key poverty predictors.

Main Methods:

• Expert ranking of survey clusters using high-resolution satellite images.
• Development of ML models using expert-defined features and a CNN using medium-resolution imagery.
• Application of ML explainability techniques to analyze feature importance and interactions.

Main Results:

• ML approaches, particularly the CNN, demonstrated higher accuracy than expert-driven methods.
• Expert-defined features showed significant overlap and limited interaction in ML models.
• The CNN approach excelled in poverty prediction using medium-resolution satellite data.

Conclusions:

• Machine learning, especially CNNs, offers superior poverty estimation capabilities compared to human expert involvement.
• ML explainability is crucial for uncovering predictive features potentially missed by human analysts.
• Integrating advanced technologies with traditional methods can enhance poverty and welfare data analysis.