Mira Kenzhebay defends her PhD thesis

Modern society depends on infrastructure systems that include critical assets such as electricity grids, transportation networks, utility corridors, urban systems, and industrial facilities. These systems face increasing pressures from climate-related hazards, ageing physical assets, rapid urban expansion, and environmental risks associated with industrial emissions. Effective monitoring therefore requires timely and spatially detailed information at scales that traditional inspection methods cannot provide consistently or cost-effectively.

Satellite Earth observation offers a scalable complement to field surveys through repeated wide-area coverage and a growing range of sensing modalities. Turning this rich data into operationally useful information, however, requires methods that can combine different data sources and remain reliable under the variability of real acquisitions.

Machine learning for satellite-based monitoring

This thesis develops and evaluates machine learning frameworks for satellite-based monitoring of infrastructure systems and associated environmental risks across three tasks: tree height estimation for vegetation encroachment analysis, urban infrastructure change detection, and methane plume detection.

The first two studies investigate multimodal fusion of optical multispectral and synthetic aperture radar (SAR) data, focusing on the extraction of spatially detailed information from complementary but heterogeneous sources. The third study addresses methane detection from hyperspectral imagery through a sensor-agnostic hybrid retrieval-learning approach designed to improve transferability across spaceborne instruments.

Across these tasks, the thesis examines how deep learning architectures can be adapted to remote sensing problems in which input modalities differ substantially in their  physical characteristics. It investigates the value of modality-specific feature processing before fusion, the potential of efficient sequence models such as Mamba for optical-SAR change detection, and the role of architectural and post-processing design in improving scene-level methane detection reliability.

The results show that operational performance depends not only on pixel-level accuracy, but also on spatial transferability, cross-sensor generalisation, and the production of usable detections under realistic deployment conditions. By addressing vegetation hazards, urban structural change, and methane emissions within a common methodological framework, the thesis demonstrates how modern machine learning can support infrastructure monitoring systems that are spatially detailed, scalable, and robust to the variability of real satellite acquisitions.

Mira Kenzhebay is an Earth Observation Data Scientist at DNV and a PhD candidate in the CI2Lab research group at Western Norway University of Applied Sciences, Norway. She is affiliated with the Department of Computer Science, Electrical Engineering and Mathematical Sciences within the Faculty of Technology, Environmental and Social Sciences. Her doctoral research focuses on artificial intelligence and multi-source satellite data analysis for infrastructure monitoring and was conducted in collaboration with StormGeo through the GridEyeS project and with DNV. She holds an MSc in Geo-Information Science and Earth Observation, specialising in Geoinformatics, from the University of Twente, and a postgraduate degree in Data Analytics from the University of Liverpool. Her professional experience centres on applying geospatial data and advanced analytics to support informed decision-making in infrastructure monitoring and environmental risk assessment.