Signatur:
S 99.0139(377)
In:
Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz Universität Hannover, Nr. 377
Materialart:
Schriftenreihen ausleihbar
Seiten:
XVI, 146 Seiten
,
Diagramme, Illustrationen, Karten
ISBN:
978-3-7696-5295-6
,
9783769652956
ISSN:
0065-5325
Serie:
Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz Universität Hannover Nr. 377
Sprache:
Englisch
,
Deutsch
Anmerkung:
Dissertation, Gottfried Wilhelm Leibniz Universität Hannover, 2021
,
Contents
1. Introduction
1.1. Motivation
1.2. Goal and Contributions
1.3. Structure of this Thesis
2. Fundamentals
2.1. Classification
2.2. Artificial Neural Network
2.2.1. Perceptron
2.2.2. Multilayer Percptrons
2.2.3. Training
2.2.3.1. Loss Function
2.2.3.2. Gradient Descent Optimization
2.2.3.3. Step Learning Policy
2.3. Convolution Neural Networks
2.3.1. Components
2.3.1.1. Convolution
2.3.1.2. Pooling
2.3.1.3. Batch Normalization
2.3.2. CNN for Image Classification
2.3.3. CNN for Semantic Segmentation
2.3.3.1. Fully Convolution Networks
2.3.3.2. U-Net
2.3.4. Training
2.3.5. Data Augmentation
3. Related Work
3.1. CNN in general
3.1.1. Image Classification
3.1.2. Semantic Segmentation
3.2. Land Cover Classification
3.3. Land Use Classification
3.3.1. Methods not based on CNN
3.3.2. CNN-based Methods
3.4. Discussion
3.4.1. Land Cover Classification
3.4.2. Land Use Classification
4. Methodology
4.1. Overview
4.2. Land Cover Classification
4.2.1. Network Architecture
4.2.2. Network Variants
4.2.2.1. Network without skip-connections
4.2.2.2. Network with elementwise addition skip-connections
4.2.2.3. Network with learnable skip-connections
4.2.3. Training
4.3. Hierarchical Land Use Classification
4.3.1. Polygon Shape Representation
4.3.2. Patch Preparation
4.3.2.1. Tiling
4.3.2.2. Scaling
4.3.2.3. Combination of tiling and scaling
4.3.3. Network Architecture
4.3.3.1. Base Network for Mask Representation: LuNet-lite
4.3.3.2. LuNet-lite with Multi-Task Learning
4.3.3.3. Achieving Consistency with the Class Hierarchy
4.3.3.4. Network Architecture for Implicit Representation
4.3.4. Training
4.3.4.1. LuNet-lite
4.3.4.2. LuNet-lite-MT
4.3.4.3. LuNet-lite-JO and LuNet-lite-BG-JO
4.3.5. Inference at Object Level
5. Datasets and Test Setup
5.1. Datasets
5.1.1. Hameln
5.1.2. Schleswig
5.1.3. Mecklenburg-Vorpommern (MV)
5.1.4. Vaihingen and Potsdam
5.2. Evaluation Metrics
5.3. Experimental Setup
5.3.1. Land Cover Classification
5.3.1.1. Test Setup
5.3.1.2. Overview of all Experiments
5.3.1.3. Prediction Variability of FuseNet-lite
5.3.1.4. Impact of the Hyperparameter Settings
5.3.1.5. Effectiveness of the learnable Skip-Connections
5.3.1.6. Performance of FuseNet-lite
5.3.1.7. Combining Datasets
5.3.2. Land Use Classification
5.3.2.1. Input Configurations
5.3.2.2. Test Setup
5.3.2.3. Overview of all Experiments
5.3.2.4. Prediction Variability of LuNet-lite-JO
5.3.2.5. Impact of the Hyperparameter Settings
5.3.2.6. Impact of Joint Optimization
5.3.2.7. Impact of the Polygon Representation
5.3.2.8. Impact of Land Cover Information
5.3.2.9. Impact of the Patch Generation
5.3.2.10. Evaluation on all Datasets
5.3.2.11. Combining Datasets
6. Experiments
6.1. Evaluation of Land Cover Classification
6.1.1. Prediction Variability of FuseNet-lite
6.1.2. Investigations of the Hyperparameter Settings
6.1.2.1. Base Learning Rate
6.1.2.2. Mini Batch Size
6.1.2.3. The Weight of the Penalty Term in the Focal Loss
6.1.3. Effectiveness of the learnable Skip-Connections
6.1.4. Evaluation on the individual Datasets
6.1.4.1. Hameln, Schleswig and MV
6.1.4.2. Vaihingen and Potsdam
6.1.4.3. Answers to the Questions raised in Section 5.3.1.6
6.1.5. Training on the combined Datasets
6.1.6. Discussion
6.2. Evaluation of Land Use Classification
6.2.1. Prediction Variability of LuNet-lite-JO
6.2.2. Investigations of the Hyperparameter Settings
6.2.2.1. Base Learning Rate
6.2.2.2. Mini Batch Size
6.2.2.3. The Weight of the Penalty Term in the Focal Loss
6.2.3. Impact of Joint Optimization
6.2.4. Impact of the Polygon Representation
6.2.5. Impact of Land Cover Information
6.2.6. Impact of the Patch Generation Approach
6.2.7. Evaluation on all Datasets
6.2.8. Training on combined Datasets
6.2.9. Discussion
7. Conclusion and Outlook
7.1. Conclusion
7.2. Outlook
References
,
Sprache der Kurzfassungen: Englisch, Deutsch
Standort:
Kompaktmagazin unten
Zweigbibliothek:
GFZ Bibliothek
Permalink