Deep learning turns satellite imagery into dynamic infrastructure map

bravo_abad · x · 2026-09-01

Researchers fine-tuned a Mask2Former model on PlanetScope imagery to map surface conditions (paved/unpaved) for 9.2 million km of global roads, achieving 95.5% coverage. The model reached 89.2% accuracy against ground truth, compared to 64.7% for existing OpenStreetMap tags. By analyzing data from 2020 and 2024, the system transforms static images into a measurement of infrastructure change, revealing a strong correlation (r=0.74) between rural road paving and human development.

Original post →

More from Research

Research channel →