Project Grant 2531102

This Project Grant award from the National Science Foundation's Geosciences Program (CFDA 47.050) will support a collaborative research project to address data irregularities in the automated classification of sea ice using artificial intelligence (AI) methods. The total funding amount is $317,922, awarded on September 1, 2025, with a project period ending on August 31, 2028.

The key objectives of this project are to develop novel, weakly supervised AI techniques that can effectively learn from existing, irregularly-labeled sea ice data to improve the accuracy of automated sea ice classification. The research will address challenges such as coarse polygon-level annotations, inaccurate labels due to human subjectivity, and inadequate pixel-level annotations needed for supervised learning. The proposed solutions include reformulating the classification problem as a multi-instance, multi-label proportion learning task, embedding confidence-aware sample selection and uncertainty quantification into the model training, and introducing a spatiotemporal multi-teacher knowledge distillation framework to leverage foundation models and task-specific models.

This research aims to advance the state-of-the-art in sea ice monitoring and mapping, which is crucial for climate monitoring, marine navigation, and offshore operations. The project will be led by the University of Colorado, a prominent public research university with extensive expertise in federal research and development across scientific domains.