Census-independent population estimation using representation learning.

Area of Science:

• Geospatial analysis
• Machine learning for population studies

Background:

• Accurate population distribution data is vital for infrastructure, resource allocation, and sustainable development goals.
• Traditional census data is infrequent and can be outdated due to migration, urbanization, and disasters.
• Existing census-independent methods often require extensive human supervision, limiting reproducibility and scalability.

Purpose of the Study:

• To explore representation learning for automated feature extraction in population estimation.
• To assess the transferability of learned representations for population mapping in Mozambique.
• To develop a more sustainable and reproducible approach to intercensal population estimation.

Main Methods:

• Applied representation learning techniques to extract features from satellite imagery.
• Utilized automated feature extraction to reduce reliance on manual annotation and public datasets.
• Assessed the transferability and accuracy of the learned representations for population estimation in Mozambique.

Main Results:

• The representation learning approach achieved population estimates comparable in accuracy to established products (GRID3, HRSL, WorldPop).
• The method demonstrated interpretability, identifying built-up areas as a key indicator of population density.
• The approach significantly reduced the need for human supervision, enhancing sustainability and reproducibility.

Conclusions:

• Representation learning offers a promising, sustainable, and reproducible method for accurate population estimation.
• Automated feature extraction via representation learning can overcome limitations of existing supervised methods.
• This approach facilitates more frequent and reliable population mapping, crucial for development initiatives.