M.Sc. (Tech) Markku Luotamo defends his doctoral thesis "Advances in Region-Based Multisource Machine Learning for Remote Sensing" on Friday the 9th of June 2023 at 12 o'clock in the University of Helsinki Physicum building, Auditorium E204 (Gustaf Hällströmin katu 2, 2nd floor). His opponent is Senior University Lecturer Jorma Laaksonen (Aalto University) and custos Associate Professor Arto Klami (University of Helsinki). The defence will be held in English.
The thesis of Markku Luotamo is a part of research done in the Department of Computer Science and in the Multi-source Probabilistic Inference group at the University of Helsinki. His supervisor has been Associate Professor Arto Klami (University of Helsinki).
Advances in Region-Based Multisource Machine Learning for Remote Sensing
Digital interpretation of remote sensing images can rely on spectral, spatial and textural image features as well as patterns over time. Machine learning has enabled progress from pixel-oriented heuristic analysis of a handful of wavelengths to automatic inference over a multitude of data sources, covering wide regions and detecting spatial texture as necessary.
Frequent scenarios and problems remain that can make digital interpretation challenging. We explore wide variations of scale in image features, annotations that are approximate, representation of subtle variation within delineated regions, sparse or pointwise observations in spatial or spectral dimensions as well as computational efficiency and operability. Also, irregularly shaped regions and sequential computational pipelines for processing large geospatial data are a common theme to the learning scenarios of this work.
This thesis makes six main contributions: 1) a multi-scale neural network architecture to segment images resource-efficiently using approximately annotated, spatially wide-reaching features such as atmospheric clouds, 2) a cloud detection method for an approximately annotated dataset with a notable accuracy improvement over a popular baseline, 3) a size-invariant, density estimate-based representation of delineated regions such as crop fields for easy multisource inference with uncertainty quantification, 4) a computational pipeline for applying inference algorithms to handheld camera hyperspectral imagery of soil samples for soil organic carbon estimation, 5) analysis of cloud detection need in a time-series inference scenario for crop yield, and 6) analysis of automation potential of possible MLOps tasks across the experiment setups.
This work uses high-resolution imagery from multispectral and synthetic aperture radar satellite instruments as well as hyperspectral images of in-situ samples. Geospatial ground truth reference data includes annotations and delineations produced by public-sector authorities as well as physical sample analysis.
Availability of the dissertation
An electronic version of the doctoral dissertation will be available on the e-thesis site of the University of Helsinki at http://urn.fi/URN:ISBN:978-951-51-9308-7.
Printed copies will be available on request from Markku Luotamo: markku.luotamo@helsinki.fi.