PhD Dissertation Proposal: Rangel Daroya, Data-Efficient Visual Learning for Scientific Measurement
Content
Speaker:
Abstract:
Computer vision models are increasingly used for scientific measurement across diverse domains. However, extracting reliable measurements remains challenging when training data is scarce and observations rely on varying or costly sensors. This thesis develops techniques for data-efficient visual learning and investigates challenges in deploying computer vision models for global-scale scientific measurement, particularly in hydrological remote sensing. To address these challenges, this thesis presents a sequence of contributions spanning task transferability, data-efficient perception, and scientific measurement. First, it develops a framework for modeling and visualizing task transferability, using box embeddings to capture asymmetric relationships across tasks. Second, it develops data-efficient perception models along two complementary axes: reducing dependence on task-specific annotations through auxiliary citizen-science supervision, and reducing auxiliary input requirements and using coarser imagery at inference through unified multitask masking and temporal super-resolution. Third, the thesis addresses the challenges of scientific measurement and deployment by evaluating physical quantities for operational applications, supporting global suspended sediment monitoring and introducing RiverScope, a multi-sensor benchmark for high-resolution river width estimation. Finally, we propose future work to model the transferability of generative models using Fisher information, and to refine river measurements from the SWOT satellite radar mission using weak supervision.
Advisor:
Subhransu Maji