One of the goals of research into remote sensing methodologies for energy and climate resource detection and monitoring is to be able to create tools that can be applied globally and nearly in real-time. Being able to keep a constant watch on ever-evolving changes to energy system infrastructure and monitor climate change’s impacts would supercharge evidence-based decision making and planning.Unfortunately, the deep learning techniques that enable analysis of satellite imagery typically require costly training data for each geographic region of interest, which makes global applications unfeasible.
Researchers at the Energy Data Analytics Lab are investigating techniques to overcome this challenge.
For example, Duke researchers are testing the use of synthetic data—artificially created images that mimic a part of the world or an object of interest—to help the deep learning techniques learn about the characteristics of a region without having actual labeled imagery from that region.
Duke researchers are also investigating use of self-supervised learning, a machine learning technique that does not require labeled imagery data. Self-supervised learning is one of the tools behind recent advances in large language models (including GPT-3, a variant of which is used in ChatGPT), but other variants are being developed for imagery data and have shown considerable promise.
With this work, Energy Data Analytics Lab researchers aim to enable upscaling of deep learning techniques for real-time, global monitoring of all forms of energy infrastructure and climate impacts.