Bhan Singh

Topic

Machine-learning Framework to Identify and Validate Biochemical Regime Clusters in the Global Blue Carbon Ecosystem

Department of Electrical and Computer Engineering

Date & location

• Tuesday, December 2, 2025

• 10:00 A.M.

• Virtual Defence

Reviewers

Supervisory Committee

• Dr. Navneet Popli, School of Electrical and Computer Engineering, UVic (co-Supervisor)

• Dr. Mihai Sima, Department of Electrical and Computer Engineering, UVic (Co-Supervisor)

External Examiner

• Dr. Daniela Constantinescu, Department of Mechanical Engineering, University of Washington 

Chair of Oral Examination

• Dr. Yu-Ting Chen, Department of Mathematics and Statistics, UVic

   

Abstract

The Earth’s climate system is undergoing profound transformation, driven by changes in natural and anthropogenic stressors that disrupt environmental balance across land, air, and sea. Among these domains, the ocean stands as both a stabilizer and a sentinel, absorbing excess heat and carbon while revealing the earliest signs of ecological stress. Yet, the ocean itself is changing, shaped by interacting forces such as temperature, salinity, oxygen depletion, depth stratification, and biological productivity. Understanding how these stressors combine to reshape marine ecosystems requires not just observation but intelligent pattern recognition. 

This thesis approaches the problem as one of learning structure within complexity. Rather than relying on political boundaries or fixed geographic regions, it asks: can we allow the data itself to define the ocean’s natural divisions? Using in-situ observations from the World Ocean Database (WOD), a machine-learning framework was developed to uncover underlying biogeochemical regimes, clusters of ocean states defined by their physical and chemical signatures. Through careful preprocessing and hierarchical spatial temporal imputation, the dataset was refined to reflect true environmental variability rather than sampling noise. 

The analysis employed multiple clustering algorithms to let ocean data “self-organize,” followed by classification models that validated and explained the separability of the dis covered regimes. This hybrid approach revealed five coherent and interpretable patterns corresponding to familiar yet dynamically interconnected oceanic systems: productive coastal upwellings, oligotrophic gyres, polar waters, oxygen-minimum zones, and transitional open-ocean regimes. Together, these patterns tell a story of a living ocean, one organized not by political maps, but by the natural language of its own chemistry and biology. 

By combining unsupervised discovery with supervised validation, the research demonstrates how global ocean observations can be transformed into quantitative, interpretable indicators of ocean health. The resulting framework contributes to the emerging vision of a digital twin ocean, a system where data, models, and machine learning work together to monitor, predict, and ultimately safeguard the resilience of the planet’s largest ecosystem.