Back to search

Harokopio University

From pixels to impact: deep learning and satellite data for rapid hazard assessment

Abstract

dc:description

Natural disasters represent one of the most severe and persistent threats to human societies and natural ecosystems. Whether driven by climatic shifts or underground geophysical dynamics, these events can unfold with devastating speed and magnitude, leading to widespread environmental damage, infrastructure loss and socio-economic disruption. In this context, timely and accurate mapping of disaster impacts is critical for effective emergency response, recovery planning and long-term risk reduction. However, traditional damage assessment methods are often slow, costly and impractical in vast or inaccessible areas. While the growing availability of satellite-based Earth Observation data offers global, continuous coverage and rich information on surface changes, the scale and complexity of these datasets pose substantial analytical and computational challenges. Deep learning has emerged as a promising solution for extracting actionable insights from large volumes of multi-temporal and multi-sensor satellite captions. Image-based architectures are particularly well suited for tasks such as classification, segmentation and change detection, assisting impact assessment both during and after the disaster. Nevertheless, training deep learning networks for hazard monitoring faces varying obstacles including limited labelled data, heterogeneous sensor characteristics, and noisy ground truth. At the same time, deployment of such models in operational contexts for supporting time-critical decision making in real-world disaster response scenarios requires robust generalization skills across regions and imaging conditions. This dissertation investigates the synergy between remote sensing and deep learning for the automated mapping of natural hazard impact, providing scientific contributions across two distinct axes. On the first axis, we present extensive manually annotated datasets in an analysis-ready format, aiming to bridge significant gaps in the literature and facilitate the training of robust large scale neural networks. We additionally conduct thorough benchmarks with a multitude of existing methods and models, setting the state-of-the-art in various disaster monitoring tasks at an unprecedented scale. On the second axis, we focus on methodological enhancements, and propose two novel deep learning architectures for burn scar mapping. Initially, a new model is designed for the delineation of burnt areas utilising high-resolution bitemporal imagery, which achieves superior results and surpasses the current state-of-the-art baseline. Subsequently, we formulate a new problem in Earth Observation by highlighting the intricacies of burn scar mapping with multi-resolution bitemporal imagery, distinguishing it from generic change detection tasks already investigated in the literature. To tackle this newly framed problem, we explore data fusion strategies motivated by the complementary nature of different sensors, then critically examine deep learning–based downscaling methods, and propose an alternative fusion approach that avoids noise imputation from intermediate resolution enhancement steps, focusing on a next-day, accurate mapping of wildfire impact. Overall, the present work advances the state of the art in disaster impact mapping exploiting both single- and multi-sensor input data for the production of accurate, trustworthy delineations. By delivering new datasets, benchmarks and custom deep learning architectures, we demonstrate how integrated remote sensing and machine learning approaches can effectively support timely decision making and climate-related risk management.

Degree

thesis:*
Grantor dc:publisher
Harokopio University
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Sdraka, Maria
  • Σδράκα, Μαρία

Subjects

dc:subject × 22

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/62034

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Sdraka, Maria; Σδράκα, Μαρία. From pixels to impact: deep learning and satellite data for rapid hazard assessment. Harokopio University, 2026. http://hdl.handle.net/10442/hedi/62034