Detecting Rural Habitat using Deep Learning on Satellite Imagery


Context and Issues

Knowing where people live and counting them is crucial in many of Masae's assignments, more specifically in rural areas (e.g., to detect underserved populations for instance). Some datasets that granularly estimate the population are already available but are often incomplete or imperfect, which was the case for Uganda.


Masae Analytics used deep learning techniques on high-resolution satellite imagery: A Convolutional Neural Network (CNN) was trained with roughly 14,000 images distributed between presence/absence of house.

The CNN yielded a 97% accuracy score in detecting human settlements, overperforming many other alternatives, especially in rural areas.