CAREER: Large-scale pretraining for Earth observation from noisy labels

Today's fleet of Earth-observing satellites offers an unprecedented view into how Earth is changing, supporting the monitoring of agriculture, infrastructure, natural disasters, air quality, and ecosystems around the world. Most applications of Earth observation, however, remain limited by the ability to make sense of the enormous volumes of satellite imagery now being collected. Modern machine learning systems require large quantities of task-specific labeled data, which are often expensive and difficult to obtain. Foundation models have transformed fields like natural language processing by learning general-purpose capabilities from massive unlabeled datasets, and similar approaches are rapidly emerging in Earth observation. As these models are increasingly applied to satellite imagery, fundamental questions remain about how well they generalize across regions, sensors, and tasks, and how their performance scales. This project aims to establish the scaling laws and generalization boundaries for Earth observation foundation models, and to develop new approaches for training and evaluating models using noisy but structured geospatial information. By reducing dependence on expert-labeled data, the project will broaden access to Earth observation technologies across scientific domains and data-scarce regions. Educational activities will train students at the intersection of machine learning and Earth observation through a new course, open-source learning materials, and undergraduate research mentoring. The research consists of three integrated components. First, the project will characterize scaling laws for Earth observation foundation models by studying how model performance varies with dataset size, model size, land cover diversity, image spatial complexity, and multi-modal satellite inputs. Second, the project will develop new pretraining strategies based on noisy but structured geospatial supervision from OpenStreetMap and other labels. Third, the project will develop a community benchmarking platform and task taxonomy for evaluating foundation models across diverse tasks, sensors, and geographic regions. Together, these activities will advance understanding of representation learning for satellite imagery while improving the scalability, robustness, and evaluation of Earth observation foundation models. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2544101 | Program: 01003031DB NSF RESEARCH & RELATED ACTIVIT,01002930DB NSF RESEARCH & RELATED ACTIVIT,01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Sherrie Wang | Institution: Massachusetts Institute of Technology, CAMBRIDGE, MA | Award Amount: $432,737 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2544101 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2544101.html