PhD Oral Exam - Chudi Wu, Civil Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Satellite-based remote sensing provides synoptic and repeatable observations for surface water monitoring, but model performance and transferability are often limited by illumination-driven variability, confounding signals, weak target separability, and poor generalization across different water bodies. This thesis aims to strengthen satellite-based surface water monitoring through three methodological responses, namely physics-guided normalization and confounder control, laboratory-informed and sensor-appropriate feature design, and rigorous evaluation of model generalization with diagnosis of transfer failures.

To reduce illumination-driven and scene-dependent variability, this thesis develops a glint-adaptive near-infrared workflow that suppresses the effects of sun glint and related confounding signals before delineation. Applied to marine and inland oil spill cases, the approach improves oil plume mapping under heterogeneous illumination conditions and demonstrates the value of physics-guided normalization, although inland transfer remains sensitive to additional confounding factors such as shallow water, complex shorelines, and residual cloud contamination.

Target separability is improved through the development of new Sentinel-2 spectral indices derived from laboratory hyperspectral reflectance data. These laboratory-informed features are evaluated for their ability to distinguish floating macroplastics from spectrally similar materials, and the results show improved classification performance with retained value across benchmark and independent datasets. Persistent confusion with other objects, however, indicates that multispectral limitations still constrain full separability in realistic scenes.

Model generalization across optically different water bodies is then examined using global and adaptive machine learning models for inland water chlorophyll-a retrieval under strict cross-lake hold-out conditions. The results show that improved within-lake performance does not guarantee reliable transfer to unseen waters. Applicability domain screening and water type classification help distinguish extrapolation from genuine cross-system shift, while further analysis identifies shifts in optically active constituent covariation and spectral ambiguity as key mechanisms of transfer failure. Overall, the thesis shows that strengthening satellite-based surface water monitoring depends not only on model choice, but also on how effectively confounding variability is constrained, how well target signals are represented within sensor limitations, and how rigorously model generalization is evaluated.