Signatur:
S 99.0139(386)
In:
Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz Universität Hannover, Nr. 386
Materialart:
Schriftenreihen ausleihbar
Seiten:
ix, 163 Seiten
,
Illustrationen, Diagramme, Karten
ISBN:
978-3-7696-5313-7
,
9783769653137
ISSN:
0174-1454
Serie:
Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz Universität Hannover Nr. 386
Sprache:
Englisch
Anmerkung:
Dissertation, Gottfried Wilhelm Leibniz Universität Hannover, 2023
,
Contents
1 Introduction
1.1 Motivation
1.2 Contributions and Scientific Goals of this Thesis
1.3 Thesis Outline
2 Basics
2.1 Machine Learning for Pixel-Wise Classification
2.2 Deep Neural Networks
2.2.1 Neuron and Multilayer Perceptron
2.2.2 Supervised Training of Neural Networks
2.2.2.1 Optimisation Strategies
2.2.3 Improving Model Generalization
2.2.4 Adversarial Training
2.3 Convolutional Neural Networks
2.3.1 Convolutional Layers
2.3.2 Pooling Layer
2.3.3 Batch Normalisation Layer
2.3.4 Activation Functions in CNNs
2.3.5 Parameter Initialisation
2.3.6 CNN Architectures
2.3.6.1 Residual Networks
2.3.6.2 Xception Network
2.4 Fully Convolutional Networks
2.4.1 Upsampling Layer and Transposed Convolutional Layer
2.4.2 Skip Connections
2.5 Appearance Adaptation
2.6 Transfer Learning and Domain Adaptation
2.6.1 Adaptive Batch Normalisation
3 Related Work
3.1 Instance Transfer
3.1.1 Explicit Instance Transfer
3.1.2 Implicit Instance Transfer
3.1.3 Hybrid Instance Transfer
3.1.4 Discussion
3.2 Representation Transfer
3.2.1 Non-adversarial Representation Transfer
3.2.2 Adversarial Representation Transfer
3.2.3 Discussion
3.3 Appearance Adaptation
3.3.1 Target-to-Source Appearance Adaptation
3.3.2 Source-to-Target Appearance Adaptation
3.3.3 Discussion
3.4 Hybrid Approaches
3.4.1 Discussion
3.5 Parameter Selection in Unsupervised Domain Adaptation
3.6 Discussion
3.6.1 Research Gap
3.6.2 Comparison to Most Similar Works
4 Methodology
4.1 Prerequisites and Assumptions
4.2 Adaptation Overview
4.3 Network Architecture
4.3.1 Classification Network C
4.3.2 Appearance Adaptation Network
4.3.3 Domain Discriminator
4.4 Training
4.4.1 Supervised Source Training
4.4.2 Joint Training for Appearance Adaptation
4.4.2.1 Joint Update of A and C
4.4.2.2 Update of D
4.5 Improving Semantic Consistency
4.5.1 Method 1: Reduction of Variability
4.5.2 Method 2: Auxiliary Generator
4.5.2.1 Architecture of G
4.5.2.2 Modifications of Adversarial Loss Terms
4.6 Entropy-based Parameter Selection
4.7 Adaptive Batch Normalization
4.8 Resolution Adaptation
5 Experimental Setup
5.1 Datasets
5.1.1 Data for Land-cover Classification using Aerial Imagery
5.1.2 Data for Bi-temporal Deforestation Detection using Satellite Imagery
5.2 Evaluation and Quality Metrics
5.3 Goals and Structure of Experiments
5.3.1 Experiment Set E1: Source Training and Na¨ıve Transfer
5.3.2 Experiment Set E2: Proposed Method for UDA
5.3.3 Experiment Set E3: Evaluation of Parameter Selection
5.3.4 Experiment Set E4: Comparison to other Strategies and Methods
5.3.4.1 Experiment Set E4.1: Comparison to other Strategies
5.3.4.2 Experiment set E4.2: Comparison to other Methods
5.3.5 Experiment set E5: Evaluation of UDA for Bi-temporal Deforestation Detection
5.4 Training Details and Hyper-parameters
5.4.1 Source Training
5.4.2 Unsupervised Domain Adaptation
5.4.3 Implementation Details of Baseline Strategies
6 Results and Discussion
6.1 Results of Experiment Set E1: Source Training and Na¨ıve Transfer
6.2 Results of Experiment Set E2: Proposed Method for UDA
6.2.1 Evaluation of Appearance Adaptation
6.2.2 Evaluation of Unsupervised Domain Adaptation
6.2.3 Combination of Appearance Adaptation with Adaptive Batch Normalisation
6.2.4 Final Comparison of Variants
6.2.5 Detailed Evaluation of Selected UDA Scenarios
6.3 Results of Experiment Set E3: Evaluation of Parameter Selection
6.4 Results of Experiment Set E4: Comparison to other Strategies and Methods
6.4.1 Experiment set E4.1: Comparison to other Strategies
6.4.2 Experiment Set E4.2: Comparison to other Methods
6.5 Results of Experiment Set E5: Evaluation of UDA for Deforestation Detection
7 Conclusions and Outlook
7.1 Conclusion
7.2 Outlook
Bibliography
Appendix
,
Sprache der Kurzfassungen: Englisch, Deutsch
Standort:
Kompaktmagazin unten
Zweigbibliothek:
GFZ Bibliothek
Permalink
