IDEAS SoCal AI & Data Science Conference 2018

Conservationists use land cover maps to quantify environmental changes like deforestation and urban development. Most land cover maps are currently made by applying rule-based classifiers to overhead imagery: this low-accuracy approach that requires manual curation, often delaying the publication of new maps by 3+ years. In this session, we will introduce the goals of Microsoft’s AI for Earth grant program and describe our work designing a high-resolution semantic segmentation model to perform land cover mapping for the Chesapeake Conservancy. Professionals in the artificial intelligence space may benefit from our discussion of lessons learned in expanding training and inference to the national scale (100+ TB imagery / 100+ K80 GPUs). Attendees will learn where to obtain our trained models, code, and methods; try out land cover mapping with our API; and apply for AI for Earth funding for their own projects.