Frontiers in AI for Human Security

Over the past five decades, AI has been one of the cornerstones of many scientific breakthroughs. With climate change remaining one of the greatest threats to humankind, we are challenged to develop new AI tools for Earth observation (EO) analysis to map the built and physical environment, monitoring, and tracking changes on the planet to support crisis management and situation awareness in humanitarian response efforts. Recent developments in large foundation models and generative AI promise new paths to transform EO data analysis. However, with over 100+ TB of Earth data collected daily, a collaborative effort is required to train effective foundation models. In this talk, I will briefly overview collaborative success stories supporting national security operational programs through AI. Next, I will review recent advances in self-supervised learning strategies and discuss challenges and opportunities, including strategies to achieve energy efficiency when training large foundation models. In conclusion, a) I will present the Quetzal foundation model for semantic segmentation and object detection with high-resolution imagery, and b) I will discuss issues with constructing global scale pretraining benchmark datasets and share new insights