Connecting the world with better maps

What this post added

This post details the application of computer vision techniques, specifically deep convolutional neural networks and weakly supervised neural networks, to analyze satellite imagery for identifying human-built structures. This analysis, combined with census data, enabled the creation of a 5-meter resolution population dataset for 20 countries, crucial for planning wireless communication technologies and backhaul networks for connecting underserved regions. The post highlights the challenges of processing billions of images and training models on unbalanced datasets, as well as the infrastructure required for large-scale analysis.