Veranstaltungsprogramm

The delineation of agricultural parcels is a pivotal task for gathering information on agricultural production systems, ensuring security of tenure and analysing ecosystems, which are strongly linked to landscape structures. Using learning techniques to automate the creation of digital field boundaries based on satellite data is a promising yet challenging task in computer vision. In this context, two specific approaches from the field of supervised learning are presented in this work. Specifically, it is analysed how the problem can be approached by leveraging traditional knowledge-based computer vision techniques embedded in a machine learning based framework in comparison to more recently evolving convolutional deep learning methods. For the former branch, a computationally lightweight edge-based, hierarchical segmentation is used, whose parametrisation is carried out in an efficient manner using Bayesian optimisation. Regarding the deep learning approach, MaskRCNN is used as a well-known baseline model and refined towards a new architecture that incorporates modular improvements including backbone attention mechanisms, rotational invariance and mask scoring. Both branches of methods are benchmarked by utilising one of the largest open-source datasets consisting of multitemporal Sentinel-2 data and correspondingly labelled ground truth for a diverse set of different agricultural production systems across Europe.

Die semantische Segmentierung von Punktwolken stellt eine herausfordernde Aufgabe dar, die in dieser Arbeit untersucht wird. Der Fokus liegt auf der Anwendung von Deep-Learning-Verfahren zur semantischen Segmentierung von photogrammetrischen Punktwolken aus Luftbildern. Es wird ein Arbeitsablauf zur Erstellung eines solchen Modells vorgestellt, welcher die Annotation einer Punktwolke mittels verschiedener Verfahren des Machine Learnings sowie die Anwendung von PointNet++ umfasst. Anhand verschiedener Experimente wird eine optimale Trainingskonfiguration ermittelt, mit dem auf einem Validierungsdatensatz eine Genauigkeit von 96,5% erreicht wird. Zusätzlich wird jedem Punkt die Wahrscheinlichkeit der Klassenzugehörigkeit zugeordnet, was eine Filterung unsicherer Punkte ermöglicht. Das Modell bietet eine geeignete Grundlage für weitere Punktwolkenverarbeitungen wie z. B. die Ableitung von digitalen Geländemodellen oder der Ableitung von Dachformen.

With the aim of extracting rotationally equivariant features, methods that achieved promising results in the field of pattern recognition were implemented and analysed to increase the tolerance towards rotated input data. Adaptions of the training pipeline, such as augmented training, as well as changes in the network architecture through the introduction of layers based on Harmonic Networks (H-Net) or Rotation Equivariant Field Networks (RotEqNet) were examined. The D2-Net architecture, which is based on the VGG16 architecture and gains state-of-the-art performance especially with regard to changes of the image domain, was used as reference. Through analyses on the HPatches dataset, significantly improved equivariance properties were achieved for all adaptation types investigated.

Baumkataster besitzen eine bedeutende Funktion für die Erfassung und Verwaltung des Baumbestands. Bisher erforderten die Aufnahme von Stammdaten der Bäume aufwendige manuelle Messverfahren. Zur Beschleunigung dieses Verfahrens wird eine Methodik vorgestellt, welche die Extraktion von Baumparametern aus Punktwolken in unterschiedlichen urbanen Umgebungen ermöglicht. Als Ergebnis des vollautomatischen Workflows liegt ein Baumkataster sowohl in tabellarischer Form als auch in Form semantischer 3D-Modelle im CityGML-Format vor. Automatisierungsbeschränkungen ergeben sich aufgrund verschiedener Punktwolkencharakteristiken, die eine Anpassung einzelner Programmparameter erfordern. Die resultierenden Baumparameter erfüllen dabei, mit Ausnahme des Brusthöhendurchmessers, die Genauigkeitsanforderungen eines Baumkatasters.

Ice shelves, the floating extensions of ice sheets, create a safety band around Antarctica by buttressing the upstream grounded ice. Loss of their stability leads to increased ice discharge contributing to sea level rise. Thus, it is important to monitor ice shelf dynamics. So far, the potential of SAR data has not yet been full exhausted as early satellites have hardly been used. To fill this research gap, this study made use of the ERS and Envisat archive within West Antarctic Pine Island Bay, a region of drastic ongoing changes. A 20-year time series (1992-2011) of ice shelf front dynamics was derived using a deep neural network architecture that combines segmentation and edge detection. The product reveals individual destabilisation patterns - most considerable for Pine Island Ice Shelf in terms of frequency of calving events and Thwaites ice tongue in terms of size of break-up.

Adequate spatial coverage of precipitation measurements is not available for large regions. Attenuation data from commercial microwave links (CML) allow precipitation estimates over existing networks, such as the cellular network. However, the processing of the data requires a distinction between wet and dry. Here, satellite data play a significant role, as they can be used in place of conventional reference data to distinguish between dry periods and precipitation events. In this work, a convolutional neural network is used to process visual and infrared cloud information from geostationary satellites to create a dry indicator. Satellite derived products exist, such as the NWC SAF products PC and PC-Ph, and are utilized as baseline products for a comparison. The evaluation shows that the developed DL product improves the performance at day and especially at nighttime. Limitations in detecting the correct rain field area is reduced by the DL product. In total the DL product improves the Matthews Correlation Coefficient value by about 0.05 compared to the PC-Ph product

For Vietnam, valuable up-to-date information on plantation forest areas and their species is not publicly available. Therefore, in this thesis, a Long Short-Term Memory (LSTM) approach for mapping rubber and acacia plantation forests was conducted for the Central Highlands of Vietnam. The input data for this time series-based classification were 12 Sentinel-2 monthly median composites for 2020. This LSTM model was compared to a random forest baseline. Accuracies were higher for the LSTM-based classification than for the random forest-based one. Also, rubber F1-values generally were higher than the F1-values for the acacia class. This can be explained by rubber being better suited for a time series classification because of its characteristic phenology compared to the evergreen acacia plant.

Variables of the soil-plant-atmosphere continuum (SPAC), like soil or plant moisture, are efficient indicators for a continuous forest monitoring under drought conditions. These variables can be estimated from active and passive microwave remote sensing to capture the drought status of temperate deciduous broadleaf forests. In this study, mainly Sentinel-1 C-band backscatter and vegetation optical depth (VOD) of AMSR2 X- and C-band as well as VOD of SMAP L-band were used from 2015 to 2020. The results reveal that X- and C-band VOD data are better suited to detect the onset of droughts affecting the canopy top (leaves), while L-band may reflect ecological memory in branches and trunks from past drought conditions. Moreover, Sentinel-1 C-band VH backscatter is very sensitive to biomass during the vegetation period.

Das Ziel dieser Arbeit ist die Analyse des Potentials multi-temporaler Daten eines terrestrischen Laserscanners zum Wachstumsmonitoring von Laubbäumen. Es wurden hauptsächlich von April bis Mai Messungen einer Amerikanischen Weiß-Eiche mit Scanintervallen von ein bis sechs Tagen durchgeführt. Für die Analyse des Wachstums werden anhand verschiedener Methoden berechnete Punktwolkenvergleiche, die Forstinventurparameter Baumhöhe, Brusthöhendurchmesser, Kronenschirmfläche, -durchmesser und -volumen, sowie aus den Punktwolken extrahierte Voxel genutzt. Mit Hilfe dieser Methoden kann aufgrund von Astabsenkungen, durch sich erhöhendes Blattgewicht, indirekt Wachstum abgeschätzt werden. Die Untersuchung der Punktanzahl der Voxel hat zu keinem Ergebnis geführt.

This work presents a novel approach to address occlusions in urban 3D data collection using LiDAR mobile mapping systems, specifically occlusions caused by parked vehicles.