Accurate, up-to-date delineation of hydrographic features is essential for hydrologic modeling, water-resource management, and disaster resilience. However, existing workflows for the U.S. National Hydrography Dataset (NHD) rely heavily on Digital Elevation Models (DEMs) and manual editing, limiting scalability, consistency, and timeliness of updates across various landscapes. This dissertation addresses the limitations of data-intensive hydrographic mapping—specifically in transferability, adaptability, and physical consistency—by developing a geospatial deep learning framework for scalable and automated hydrographic delineation through three interconnected studies.
The first study addresses the limitation of transferability by leveraging transfer learning with convolutional neural networks pre-trained on ImageNet to improve streamline segmentation accuracy and generalization across heterogeneous landscapes. The second study tackles the challenge of adaptability through Model-Agnostic Meta-Learning (MAML), enabling models to rapidly adjust to new geographic regions with minimal labeled data using Alaska’s IfSAR dataset.
Building upon these advances, the final study addresses the issue of physical consistency by developing a multimodal, multitask deep learning framework that fuses DEM, synthetic aperture radar (SAR), optical, thermal, and learned embeddings to jointly predict streamlines and flow directions. This framework integrates complementary information sources to enforce topological coherence while enhancing scalability and robustness for continental-scale hydrographic mapping.
Taken together, this dissertation advances an integrative, cyberGIS-enabled deep learning methodology that enhances the automation, transferability, and scalability of national hydrography production – laying the foundation for next-generation hydrographic data workflows and supporting intelligent environmental management.