Dissertations and Theses

Supervising

Dissertations and Theses

PhD Dissertations

Master Theses

Bachelor Theses

PhD Dissertations

Search by topic

Thank you! Your submission has been received!

Oops! Something went wrong while submitting the form.

2023

Scalable Deep Learning for Remote Sensing with High Performance Computing

Author:

Rocco Sedona

Supervisors:

Morris Riedel, Gabriele Cavallaro, Matthias Book

University of Iceland

Advances in remote sensing (RS) missions in recent decades have greatly increased the volume of data that is continually acquired and made available to end users, who can utilize it in a variety of Earth observation (EO) applications. land cover (LC) maps play a key role in monitoring the Earth’s surface, providing scientists and policymakers with an accurate view of the evolution of the landscape and helping them address pressing questions, from efficient resource planning to resilience to climate change. Due to the use of classical machine learning (ML) and more recently of deep learning (DL) methods, the information content of RS data can be exploited to an unprecedented degree, fostering research, development, and deployment of workloads to address open challenges for EO applications, including LC classification. However, the larger size of the datasets needed to train state-of-the-art (SotA) DL models and the need to utilise them at scale increases the time to deployment, which can hinder their effective utilization. Adopting strategies for distributed deep learning (DDL) on high performance computing (HPC) systems provides the opportunity to speed up the training of the models, allowing faster development times for researchers. Since space agencies operate a variety of missions, data acquired by different sensors can be used to increase the temporal resolution at which a certain area is observed, with potential improvements in the accuracy of the ML/DL models. The thesis objectives are formulated with these premises in mind and were investigated using a combination of methodologies to exploit the dedicated resources of HPC systems, contributing to addressing new questions on the adoption of DDL methods for EO applications and to familiarize the RS community with such approaches, which can be of great value.

‍

View the recording PhD defence

‍

To the thesis

Master Theses

Search by topic

Thank you! Your submission has been received!

Oops! Something went wrong while submitting the form.

2023

Hyperspectral Image Segmentation using Quantum Annealing

Author:

Haytam El Merabeti

Supervisors:

Gabriele Cavallaro and Piotr Gawron

Universidade da Coruña

Hyperspectral imaging plays a crucial role in various fields, such as remote sensing, med- ical imaging and agriculture, due to its ability to capture rich spectral information. However, the analysis and segmentation of hyperspectral images present significant computational challenges, mainly attributed to the high dimensionality and complex spectral characteris- tics of the data. This master's thesis project, aims to explore the potential of quantum annealing as a novel approach for hyperspectral image segmentation.

‍

To the thesis

2023

Automatic Workflow Based on Unsupervised Machine Learning Methods that Leverages Satellite and Geo-tagged Social Media Data

Author:

Savyasachi Pandey

Supervisors:

Benjamin Berkels, Gabriele Cavallaro

RWTH Aachen University

To the thesis

2023

A Scalable Processing Pipeline Based On Deep Neural Network Models For Transfer Learning In Earth Observation Applications

Author:

Shivangi Kirti

Supervisors:

Benjamin Berkels, Gabriele Cavallaro

RWTH Aachen University

To the thesis

2022

Scaling the Production of Land-Cover Maps with an Automatic HPC Workflow based on Machine Learning

Author:

Liang Tian

Supervisors:

Gabriele Cavallaro, Morris Riedel, Klaus Diepold

Technical University of Munich

Remote sensing data are nowadays available on different scales. One of the most important applications of these data is to generate thematic products such as land cover maps. Land cover maps represent the spatial information of physical cover-age of the Earth’s surface. However, the land cover maps are normally released every few years. During this period, a lot of changes could happen. On the other hand, generating land cover maps at a global scale is time-consuming and requires much computing power.

This thesis extends a workflow that can automatically update the land cover maps based on free remote sensing images and machine learning algorithms, once the data is accessible. First, a small area of the Netherlands is selected as a prototype. After it is implemented and validated, it is ported to high performance computing system and scaled to a larger area, the whole country of Netherlands. It delivers a convincing result and the generated land cover maps could be then applied for other purposes.

The whole workflow is scalable, unsupervised, and modular. Combined with the high performance computing system of Jülich Supercomputing Center, it greatly reduces the time of generating land cover maps and increases the frequency of the update.

‍

To the thesis

2022

Leveraging Quantum Annealing for the Development of a Multiclass Support Vector Machine

Author:

Amer Delilbasic

Supervisors:

Gabriele Cavallaro, Farid Melgani

University of Trento

Quantum Annealing (QA) is a quantum computing metaheuristic aimed at solving optimization problems. It has been successfully applied to a variety of machine learning tasks. This work presents an analysis of the current state-of-the-art and the related topics, with a focus on the Support Vector Machine (SVM) algorithm and QA. A novel QA-based SVM formulation for multiclass classification based on a single optimization step is proposed. The algorithm is then validated on the D-Wave Advantage quantum annealer. The obtained results have evinced good overall results, comparable to classical algorithms for classification.

‍

Get thesis

To the thesis

2021

Support Vector Regression With D-wave Quantum Annealer

Author:

Edoardo Pasetto

Supervisors:

Gabriele Cavallaro, Farid Melgani

University of Trento

The objective of this this master’s thesis is to propose a method to optimize the cost function re-lated to the training phase of a Support Vector Machine (SVM) for regression. The novelty in the optimization process that was developed in this study consists in the usage of Quantum Annealing (QA) for the minimization of the cost function. The hardware used for the annealing process is the D-wave Advantage system that was released in 2020. The implementation proposed in this work, that will be referred to as Quantum Support Vector Regression (QSVR), was tested on simple test cases of one variable functions in order to carry out a preliminary assessment on toy examples. The functions considered in the experimental analysis were a sine function, a sinc function and a triangular function. The number of training samples was kept low in each experiment beacuase the objective was to study the model in simple cases. Moreover, the hyperparameters were chosen in such a way to favor overﬁtting. In the optimization process the annealer provides different solutions and therefore 5 different methods to combine such solutions were proposed and then compared with the traditional SVR as a reference in the experiments. One interesting fact that emerged during the experimentation is that the different methods for combining the solutions performed very differently depending on the function considered for the testing. This suggests that there is not a best method overall but it should instead be chosen speciﬁcally based on the test case. Further research in the future will study the generalization capabilities of the QSVR and will also consider more complex functions and test cases for the training and the testing phase.

‍

Get thesis

To the thesis

2020

Super-Resolution of Sentinel-2 Images with Generative Adversarial Networks

Author:

Run Zhang

Supervisors:

Bastian Leibe, Morris Riedel, Gabriele Cavallaro, Jenia Jitsev

RWTH Aachen University

Remote Sensing (RS) image sensors onboard satellites and airborne platforms are required to provide highquality Earth observation data. However, due to the limitation of sensor resolution, satellite orbital altitudes, spaceground communication bandwidth, etc., the raw acquisitions are often of low resolution (LR) and low signal to noise ratio (SNR). One important indicator to measure the quality of a RS image is ground sampling distance (GSD), i.e., the distance between two consecutive pixel centers measured on the ground. Small GSD means a finer representation with more useful information of the earth surface. Operational satellites can not meet the growing GSD requirements of many new generation scientific applications. Therefore, it is urgent to propose a post-correction method to enhance the GSD of raw RS acquisitions, or improve their spatial resolution. This thesis proposes a scalable approach based on deep learning for the super-resolution of multi-resolution multispectral RS images.

Get thesis

To the thesis

2019

Classification of Polar Stratospheric Clouds with Machine Learning Methods

Author:

Rocco Sedona

Supervisors:

Farid Melgani, Gabriele Cavallaro, Lars Hoffmann, Reinhold Spang

University of Trento

Abstract

‍

To the thesis

2019

Translation of Remote Sensing Image for the Classification of Unlabeled SAR Data with Deep Convolutional Generative Adversarial Networks

Author:

Julius Lange

Supervisors:

Gabriele Cavallaro, Morris Riedel, Heinz-Wilhelm Hübers

Humboldt University of Berlin

Remote Sensing (RS) involves the contact-free measurement and processing of data about the Earth's surface. Machine learning, specifically deep learning, has been increasingly used in RS to analyse data and improve results. The aim of this thesis is to train a land cover classification algorithm for RS using an unsupervised learning approach. The method involves training an image translator algorithm using a labeled optical dataset and an open dataset of co-registered optical and synthetic aperture radar (SAR) images, generating an artificial SAR dataset with labels, and using this dataset to train a SAR classifier in a supervised fashion. The image translator is based on generative adversarial networks (GANs) and the SAR classifier uses a convolutional neural network (CNN).

Get thesis

To the thesis

2018

Brain Image Classification with Support Vector Machines using Self-Dual Attribute Profiles

Author:

Steinunn Gróa Sigurðardóttir

Supervisors:

Gabriele Cavallaro, Morris Riedel

University of Iceland

A more detailed knowledge of the human brain is desirable. Automating the process of reviewing brain images and detecting abnormalities of those would be a significant effort in that process. In this research, a labelled data set of high resolution brain slice images is analysed. Each pixel of the images has been manually labelled, depending on which area of the brain the pixel belongs to. The aim of this research is to perform pixel-wise classification, utilizing the data set of brain images. The objective is to recognise cortex regions from other areas of the brain. The challenge of the work lies in the magnitude of an image, that is approximately 37.000.000 pixels per image. The volume of the images introduce computational complexity in terms of memory and processing time, which encourages the employ of a High Performance Computing (HPC) system. The proposed approach is to perform feature extraction with Self-Dual Attribute Profiles (SDAP) and to use Support Vector Machines (SVM) for the classification. The obtained classification results indicate that Support Vector Machines are suitable for classification of brain images.

Get thesis

To the thesis