Abstract
dc:descriptionThe rich and multi-source information provided by Earth Observation (EO) instruments creates new possibilities for studying the complex processes that govern the Earth System. This potential is further amplified by the recent surge of freely available multi-sensor EO data, which provides fertile ground for integrating EO with deep learning, driving a shift toward data-driven monitoring and forecasting. This emerging synergy holds the potential to advance a wide range of highly critical applications with direct social impact, such as volcanic activity monitoring, rapid flood mapping, and wildfire risk estimation. Deep learning has already demonstrated remarkable success in various domains, including Computer Vision and Natural Language Processing. This progress has been largely driven by the introduction of large-scale annotated datasets like ImageNet, which enable the learning of rich and meaningful representations, significantly enhancing predictive performance. However, extending deep learning methods to EO is far from straightforward, as the domain introduces its own distinct challenges. Curating reliable EO datasets is costly and labor-intensive, often requiring expert annotations, while remaining vulnerable to label noise. Moreover, EO imagery differs fundamentally from natural images: it is highly multimodal, spans varying spatiotemporal scales, and exhibits variations in object size, resolution, acquisition modes, sensor types and environmental conditions. In contrast to the natural image domain, every element of an EO scene carries potential relevance, eliminating the notion of background and making class definitions use-case dependent. As a result, creating a well-curated dataset for each task is infeasible due to the ever-increasing set of possible applications and sensor combinations. In addition, the critical nature of EO applications requires models that are not only accurate but also trustworthy, further complicating the adoption of deep learning methods. This thesis addresses these challenges by advancing deep representation learning for EO along three main directions: a) developing task-specific EO models and datasets; b) designing task- and sensor-agnostic representation learning methods moving towards unified EO foundation models; and c) investigating the generalization capacity of pretrained representation uncertainty in EO, working towards zero-shot data uncertainty estimation across the EO domain and ultimately contributing to the development of trustworthy deep learning models. Together, these contributions advance the current state of representation learning for EO data both in terms of performance and trustworthiness.
Degree
thesis:*- Grantor dc:publisher
- Harokopio University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Bountos, Nikolaos-Ioannis
- Μπούντος, Νικόλαος-Ιωάννης
Subjects
dc:subject × 16- Μηχανική μάθηση
- Υπολογιστική όραση
- Παρατήρηση γης
- Βαθιά μάθηση
- Machine learning
- Computer vision
- Earth observation
- Deep learning
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Επιστήμη ηλεκτρονικών υπολογιστών
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
- Computer Science
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60956