Hydrology-Aware Deep Learning for Physically Consistent Wetland Classification

Type of project

M.Sc.

Contact

Puzhao Zhang (puzh@dhigroup.com)

Topic

    Wetlands are critical ecological buffers, yet they remain one of the most difficult land-cover types to map accurately due to high seasonal variability and extreme spectral similarity to upland land-cover. This thesis explores a novel approach using Alpha Earth Foundation Embeddings—highly compressed, 64-dimensional numerical representations of Earth's surface that capture long-term vegetation patterns and moisture cycles. However, when spectral features and topographic features are stacked directly, standard deep learning models are often "lazy," frequently overfitting to simple elevation data. This leads to physically inconsistent results and poor generalization across varying landscapes.

The core objective of this thesis is to investigate advanced spectra-topo fusion techniques that move beyond simple feature stacking. Rather than treating topography as just another input channel, this research proposes novel architectures that use topographical features—specifically Height Above Nearest Drainage (HAND), slope etc.)—as physical constraints to dynamically gate the Alpha Earth embeddings. In this framework, topography generates an attention mask that identifies hydrologically favorable zones. The model is thus trained to prioritize complex spectral-temporal analysis only within these regions. Conversely, in hydrologically unfavorable areas where wetland formation is physically impossible, the topographic constraint automatically filters out false positives, preventing the model from being misled by spectral mimics.

*The expected outcome is to develop a "hydrologically intelligent" segmentation model that is not only more accurate but truly generalizable across diverse terrains, from coastal lowlands to inland uplands. *

This project is tailored for students with a strong background in large data processing, Earth Observation (ideally Google Earth Engine) and Machine Learning (particularly PyTorch), and experience in GPU environments is a plus.

Relevance to DHI

This thesis directly supports DHI’s research on wetlands by addressing a core challenge: producing globally consistent, physically meaningful wetland maps that generalize across regions. The outcome is a more robust and explainable wetland mapping approach that can be integrated into DHI’s operational, global‑scale Earth Observation workflows.

What is the student's contribution

The student will contribute methodologically by developing and testing hydrology‑aware deep learning architectures that use topographic variables (e.g. HAND, slope, TWI) as physical constraints to guide and regularize large‑scale spectral‑temporal embeddings, reducing false positives and improving generalization.