Increasing the Confidence of Predictive Uncertainty: Earth Observations and Deep Learning for Poverty Estimation

• Published in: IEEE transactions on geoscience and remote sensing 2024, Vol.62, p.1-13
• Main Authors: Kakooei, Mohammad, Daoud, Adel
• Format: Article
• Language: English
• Subjects: Accuracy; Data models; Deep learning; Deep learning(DL); Earth; Estimation; Indexes; Landsat; Machine learning; Nighttime light; nighttime light (NL); Observational learning; Poverty; Poverty estimation; Prediction uncertainty; Predictions; Predictive models; Risk management; Satellite imagery; Sustainable development; Sustainable Development Goals; Training; Uncertainty
• ISSN: 0196-2892; 1558-0644; 1558-0644
• DOI: 10.1109/TGRS.2024.3392605

Reducing global poverty, particularly in low- and middle-income countries, is a critical objective of the sustainable development goals (SDGs). To track progress toward these goals, high-frequency, granular geo-temporal data that capture changes at the neighborhood level is essential for researchers and policymakers. Recent advancements in methodology have combined machine learning (ML) and Earth observations (EOs) for poverty estimation, thereby addressing significant data gaps. However, a notable limitation of these EO-ML methods is their frequent deployment without a robust mechanism to quantify predictive uncertainty. Understanding this uncertainty is crucial for making informed decisions, effectively managing risks, and instilling confidence in users and stakeholders regarding the model's predictions. Although deep learning (DL) offers methods to quantify predictive uncertainties, their reliability is often constrained, failing to accurately reflect the underlying variations in predictions. Our proposed method aims to enhance confidence in predictive uncertainty without sacrificing accuracy. It begins by integrating an external model to explicitly capture data variability. Subsequently, we employ two orthogonal metrics-accuracy and uncertainty-to evaluate the influence of training data, especially in scenarios involving satellite imagery (e.g., selecting a subset of source domain countries for prediction in the target country). By applying these metrics, we formulate criteria to assess the importance of choosing specific countries from the source domain as training data. Our analysis highlights the effectiveness of this methodology in situations where the target country offers high-dimensional data, like satellite images, but faces a shortage of adequate training samples for DL models.