Theses
Bachelor
STML
Developing a web-based forest-fire spread model for Germany
With climate change more frequent and more intense forest fires pose a serious threat to the environment and societies. In a collaboration between the Institute of Landscape Ecology and the Institute of Geoinformatics we aim to extend a web-based burn simulator (known as Ember-sim) to simulate the spread of fire across the landscape based on established fire behaviour models. The burn simulator will then be used for educational purposes, research, and training professional fire practitioners, governmental officers and volunteers from the community.
The project can be delivered at the BSc or MSc level and the potential candidate will be in charge of extending Ember-sim originally developed for the Australian continent, for Germany. Project starting date is open until position is filled.
Are you experienced with JavaScript and have interest in climate change-related topics?
Please contact:
Prof. Mana Gharun mana.gharun@uni-muenster.de
Or
Dr Christian Knoth christian.knoth@uni-muenster.de
Contact: Christian Knoth
Author: Ahmed Aly
Supervisor: Mana Gharun, Christian Knoth
SILMechanismen zur Erstellung von Anfahrtsskizzen
Anfahrtskizzen sind Karten, die optimiert sind, um die Route aus verschiedenen Richtungen zu einem bestimmten Zielort zu zeigen. Zumeist sind sie manuell erstellt, wobei
- Informationen selektiv ausgewählt wurden
- schematisiert dargestellt wurden
Die Arbeit soll Anfahrtsskizzen analysieren und die dahinterliegenden Mechanismen extrahieren, wie Anfahrtsskizzen erfolgreich visualisiert werden können. Die Ergebnisse werden systematisch in einer Nutzerstudie getestet.
Beispiele für Anfahrtsskizzen:
http://www.hautarztpraxis-muenster.de
Contact: Angela Schwering
Author: Jens Golze
Supervisor: Angela Schwering
Co-supervisor: Vanessa Joy Anacta
Earth Observation (EO) data cubes are multidimensional arrays of spatial and temporal data, crucial for monitoring and analyzing environmental changes. Machine learning (ML) models applied to EO data cubes enable advanced data analysis and predictive capabilities. However, the diversity of programming languages used in the spatial data science and geoinformatics community, particularly R and Python, poses challenges for interoperability and reproducibility of these ML models.
The outcomes of this research are expected to facilitate smoother integration and collaboration among spatial data scientists and geoinformatics professionals who rely on different programming environments, promoting the reproducibility and interoperability of EO data analysis projects. This work will contribute to the broader goal of advancing geospatial data science by bridging the gap between diverse computational ecosystems.
Use Case:
Carrying out spatial-temporal analysis, such as time-series crop classification in Germany, leveraging the ONNX interoperability format : https://onnx.ai/
- Model Portability: How can a deterministic machine learning model, such as Support Vector Machine (SVM), be trained on Earth Observation data cubes in Python and then ported to R using the ONNX format?
- Performance Evaluation: What are the differences in performance and accuracy of the SVM model when ported from Python to R for time-series crop classification in Germany
- Interoperability Challenges: What are the challenges and potential solutions in ensuring interoperability and reproducibility of machine learning models between Python and R programming environments using ONNX?
- What is the feasibility of implementing identical deep learning models for Earth Observation data cubes in R, Python, and Julia, and ensuring their interoperability?
- How do the available tools and libraries for machine learning in R, Python, and Julia compare in terms of ease of use, performance, and integration with EO data cubes?
- What are the differences in command structure and interface among R, Python, and Julia for machine learning tasks related to EO data cubes, and how do these differences impact the reproducibility and interoperability of the models?
Please contact:
Brian Pondi brian.pondi@uni-muenster.de
and
Edzer Pebesma edzer.pebesma@uni-muenster.de
Contact: Brian Pondi
Author: Jonas Starke, Sanju Shree Kumar
Supervisor: Edzer Pebesma
Co-supervisor: Brian Pondi
SITCOMApps zur Förderung nachhaltigen Verhaltens
Eine Vielzahl verfügbarer Apps soll ihre Nutzer*innen dabei unterstützen, sich auf nachhaltigere Art zu verhalten. Dies kann zum Beispiel das Mobilitätsverhalten betreffen (Rad statt Auto), den Energiekonsum (Pulli statt Heizung) oder auch Aktivitäten (Mitwirkung bei Urban Gardening).
Ziel dieser Bachelorarbeit ist es, eine repräsentative Auswahl dieser Apps systematisch zu analysieren, um einen Überblick über die verschiedenen Arten von Apps, der Hauptthemen und -strategien zu gewinnen. Von Interesse sind insbesondere auch welche Techniken aus dem Bereich "Nudging" und "Persuasive Technology" zum Einsatz kommen.
Das Thema kann auf unterschiedliche Weise bearbeitet werden, je nach Wunsch der bearbeitenden Person, z.B. als reine Analyse basierend auf detaillierter Betrachtung der Apps, als Nutzerstudie oder einer Kombination aus beidem. Denkbar ist auch der Entwurf einer App als Mockup/Prototyp, die bestehende Schwächen von verfügbaren Apps überwindet.
Contact: Christian Kray
Author: Felix Disselkamp
Supervisor: Chris Kray
Co-supervisor: Simge Özdal-Oktay
Master
External“Feature Info” – Visualizing GIS Background Information
Contact: Morin Ostkamp
Author: Yevgeniya Litvinova
Supervisor: Christian Kray
Co-supervisor: Morin Ostkamp
STMLDeveloping a web-based forest-fire spread model for Germany
With climate change more frequent and more intense forest fires pose a serious threat to the environment and societies. In a collaboration between the Institute of Landscape Ecology and the Institute of Geoinformatics we aim to extend a web-based burn simulator (known as Ember-sim) to simulate the spread of fire across the landscape based on established fire behaviour models. The burn simulator will then be used for educational purposes, research, and training professional fire practitioners, governmental officers and volunteers from the community.
The project can be delivered at the BSc or MSc level and the potential candidate will be in charge of extending Ember-sim originally developed for the Australian continent, for Germany. Project starting date is open until position is filled.
Are you experienced with JavaScript and have interest in climate change-related topics?
Please contact:
Prof. Mana Gharun mana.gharun@uni-muenster.de
Or
Dr Christian Knoth christian.knoth@uni-muenster.de
Contact: Christian Knoth
Author: Ahmed Aly
Supervisor: Mana Gharun, Christian Knoth
SILCognitively plausible model of place
This idea behind this comes from Luescher and Weibel (2010) - "we might for example present a different answer to the question ‘shopping opportunities in the city centre?’ to elderly people than to young people (the former avoiding the night club district because they might perceive it unsafe)."
Despite the notion of place being widespread in natural language, it is still difficult to model it in a GIS (which is more concerned with location). The student can identify a suitable site, and attempt to study how the notion of place differs among participants of different age groups (young vs. old), residence (e.g living in the city vs. suburbs) etc. The analysis could offer insights into whether there are differences in the way a place is perceived based on these factors and has potential implications for contextualized location based services.
Contact: Imad Humayun
Author: Tobias Tresselt
Supervisor: Angela Schwering
Co-supervisor: Imad Humayun
SILCreating schematic maps in mobile devices
Schematic maps are maps that intentionally distort and misrepresent geometries in order to simplify maps. Oftentimes, schematic maps give a better overview of the environment, because they focus only on important aspects. The displays of mobile devices offer only limited space. This thesis topic aims at the development of an algorithm to schematize maps automatically for mobile devices.
Schematization refers to misrepresentations of shape and size of spatial objects and distortions of distance and direction relations among them. The research starts from findings by Latecki and Lakämper, who developed the discrete curve evolution, an algorithm to schematize topographic maps. Barkowski et al. applied this algorithm to map schematization.
References:
Latecki, L.J. and R. Lakämper (2000). Shape similarity measure based on correspondence of visual parts. IEEE Trans. Pattern Analysis and Machine Intelligence (PAMI) 22(10): 1185-1190.
Barkowsky, T., L.J. Latecki, and K.-F. Richter (2000). Schematizing maps: Simplification of geographic shape by discrete curve evolution. Spatial cognition II: Integrating abstract theories, empirical studies, formal methods and practical applications. C. Freksa, et al., Springer: 41–53.
Contact: Angela Schwering
Author: Andreas Ohrem
Supervisor: Angela Schwering
SILVisualization of global landmarks on mobile devices.
In the WayTo research project, we developed a prototype to visualize global landmarks which are outside the map section displayed on a mobile device by visualizing them at the edge of screen to indicate directions.
An amendment to this design is to incorporate distance into the design of icons. The thesis could explore different techniques (e.g. halo (circles at the edge), wedge (triangle at the edge), or distance-encoding arrows). These methods distinguish only two or three levels (or four maximum) of icons size (that means we only have 4 different sizes of each global landmarks to indicate very close, close, far, and very far) on screen. Although there is hardly any empirical data to support that, drawing from cartographic theories regarding visual variables in animated map making, humans are not able to distinguish changes on map representation at fine levels.
Possible bachelor and master thesis could address:
1. Implementing distance and direction indicating landmarks (DDL) display on mobile phones
2. Investigating the effects of DDL on mobile screens on spatial orientation and spatial knowledge
3. Comparing the variations of DDL design (gradient to true distance vs. categorical distance)
4. Comparing the effects of DDL with Halo and Wedge design
Reference: Li, R.; Korda, A.; Radtke, M.; Schwering, A. (2014): Visualising distant off-screen landmarks on mobile devices to support spatial orientation. https://www.uni-muenster.de/forschungaz/publication/99385?lang=en
Contact: Angela Schwering
Author: x
Supervisor: Angela Schwering
SILEnhancing Dynamic Geometry with Artificial Intelligence: Applications in Medicine and GIS
Powerful methods in an area called "dynamic geometry" provide a rapid way of solving a wide range of spatial reasoning problems. For example, as a user manipulates some shapes (circles, lines, points) in an "intelligent" sketch, the sketch automatically updates itself to maintain certain spatial constraints (e.g. lines being tangent to circles, lines being parallel to each other, and so on).
In medicine these shapes could represent different types of cells that have been automatically recognised from images of a tissue section, and the task could be to determine whether the cells are cancerous (histopathology). In geographic information systems these shapes could represent streets, buildings, and landmarks recognised from satellite images or directly from a hand-drawn sketch.
In this thesis project you will be:
- extending the spatial language of dynamic geometry to work with common-sense qualitative spatial relations (near, left of, inside, etc.)
- integrating dynamic geometry into a knowledge representation (artificial intelligence) framework
You will then use this enhanced technology for a range of exciting tasks, such as:
- improving computer-based image recognition by automatically correcting errors based on knowledge about objects in the domain
- providing new ways of interacting with complex data using "intelligent" sketches
Through this research project you will develop skills and experience in the application of methods in artificial intelligence (AI). You will be introduced to the necessary tools and existing projects to build on. No prior experience with methods in AI is necessary (you will be given considerable support in this area).
Contact: Carl Schultz
Author: Nikolai Gorte
Supervisor: Angela Schwering
Co-supervisor: Carl Schultz
SILEmbodied 3d Isovist as a Predictor of Spatial Experience
Isovist analysis is used to predict how people behave in and 'feel about' the geometry of space. One of the best-documented examples of such relation is provided by Wiener et al. (2007). In this work, the authors controlled the shape of a Virtual Reality environment and measured how people experience individual spaces, depending on its underlying isovist properties.
Recently, 3d isovist analysis became increasingly popular in contexts where its traditional 2d counterpart suffers major limitations. However, it is not clear if the influence of isovists on human experience can be directly extrapolated from the 2d to the 3d analysis. In this thesis, the student will replicate the VR experiment of Wiener and colleagues, taking into account the vertical component of human experience and using a 3-dimensional isovist in the analysis of final results.
Wiener, J. M., Franz, G., Rossmanith, N., Reichelt, A., Mallot, H. A., & Bülthoff, H. H. (2007). Isovist analysis captures properties of space relevant for locomotion and experience. Perception, 36(7), 1066 – 1083. http://doi.org/10.1068/p5587
Krukar, J., Schultz, C., Bhatt, M., (forthcoming). Towards Embodied 3d Isovists: Incorporating cognitively-motivated semantics of `space’ and the architectural environment in 3D visibility analysis
Contact: Jakub Krukar
Author: Charu Manivannan
Supervisor: Jakub Krukar
SILAn Artificial Intelligence System that Learns by Spatial Induction
"Learning by induction" is the ability to take a number of observations or examples and discover rules that account for the observations - it's the ability to generalise from examples.
Being able to generalise spatial patterns from observations is essential for artificial intelligence systems to perform a variety of tasks in a flexible and robust way. We don't want to tell the computer system exactly how to solve every specific problem it will encounter. Instead, we want to give it some examples and have the system use common sense and background knowledge to figure out ways of solving new problems that have a similar structure.
In this thesis project you will be:
(a) integrating a new powerful "common sense" spatial reasoning method called Enhanced Geometric Constraint Solving within a more general artificial intelligence framework for inductive spatial learning;
(b) evaluating the system in a range of exciting applications in geographic information science, architectural design, cognitive psychology (analogical reasoning), and medicine (histopathology).
Through this project you will develop skills and experience in the application of methods in artificial intelligence (AI). You will be introduced to the necessary tools and existing projects to build on. No prior experience with methods in AI is necessary (you will be given considerable support in this area).
Contact: Carl Peter Leslie Schultz
Author: Lars Syfuß
Supervisor: Edzer Pebesma
Co-supervisor: Carl Schultz
Earth Observation (EO) data cubes are multidimensional arrays of spatial and temporal data, crucial for monitoring and analyzing environmental changes. Machine learning (ML) models applied to EO data cubes enable advanced data analysis and predictive capabilities. However, the diversity of programming languages used in the spatial data science and geoinformatics community, particularly R and Python, poses challenges for interoperability and reproducibility of these ML models.
The outcomes of this research are expected to facilitate smoother integration and collaboration among spatial data scientists and geoinformatics professionals who rely on different programming environments, promoting the reproducibility and interoperability of EO data analysis projects. This work will contribute to the broader goal of advancing geospatial data science by bridging the gap between diverse computational ecosystems.
Use Case:
Carrying out spatial-temporal analysis, such as time-series crop classification in Germany, leveraging the ONNX interoperability format : https://onnx.ai/
- Model Portability: How can a deterministic machine learning model, such as Support Vector Machine (SVM), be trained on Earth Observation data cubes in Python and then ported to R using the ONNX format?
- Performance Evaluation: What are the differences in performance and accuracy of the SVM model when ported from Python to R for time-series crop classification in Germany
- Interoperability Challenges: What are the challenges and potential solutions in ensuring interoperability and reproducibility of machine learning models between Python and R programming environments using ONNX?
- What is the feasibility of implementing identical deep learning models for Earth Observation data cubes in R, Python, and Julia, and ensuring their interoperability?
- How do the available tools and libraries for machine learning in R, Python, and Julia compare in terms of ease of use, performance, and integration with EO data cubes?
- What are the differences in command structure and interface among R, Python, and Julia for machine learning tasks related to EO data cubes, and how do these differences impact the reproducibility and interoperability of the models?
Please contact:
Brian Pondi brian.pondi@uni-muenster.de
and
Edzer Pebesma edzer.pebesma@uni-muenster.de
Contact: Brian Pondi
Author: Jonas Starke, Sanju Shree Kumar
Supervisor: Edzer Pebesma
Co-supervisor: Brian Pondi
SILValidity of sketch maps under varied tasks
Blades, M., 1990. The reliability of data collected from sketch maps. Journal of Environmental Psychology, 10(4), pp.327–339.
Contact: Jakub Krukar
Author: Antonia van Eek
Supervisor: Angela Schwering
Co-supervisor: Jakub Krukar
ERSLComparison of spatial validation strategies in machine learning for environmental science
A major task in environmental science is to obtain spatially comprehensive data from limited field samples (e.g. climate stations, soil profiles, vegetation records,...). This is often done using machine learning algorithms that learn the relationships between field data and remotely sensed predictor variables (e.g. from satellites). The developed model is then used to make spatial predictions for the entire area of interest (i.e. create a "map" of the variable of interest).
Such a map is only valuable when the error of the model is known. The error assessment however causes major difficulties for models with spatial dependencies and cause standard validation methods to fail. There is increasing consent in literature that spatial validation is necessary and several strategies have been proposed (e.g. Roberts 2018, Meyer 2018, Valavi 2018, Brenning 2012). However, little research is done on how different strategies compare, which however is important to know for model comparisons and for finding the "right" validation strategy for a dataset.
This project aims at comparing and (as far as possible) evaluating different validation strategies for machine learning based spatial mapping of environmental variables. Usually in most projects the "true" performance is hard to assess due to the fact that the reference data are limited. Therefore we want to adress the problem using reference data that are available in a spatially continuous way hence providing a continuous reference. Such data are rare but a possible research task would be available in the field of rainfall monitoring: The RADOLAN dataset contains high quality rainfall data from a radar network in 1km spatial resolution and serves as an excellent reference set. We then want to model rainfall for Germany using data from the the geostationary satellite sensor MSG SEVIRI and auxiliary predictor variables such as elevation (see Kühnlein 2014 for the idea of mdoelling rainfall based on MSG SEVIRI).
Using the RADOLAN data we simulate different sampling designs (hence we simulate raingauges in different constellations: random, clustered,...) and train machine learning models to predict rainfall from MSG SEVIRI in a spatial way. Different strategies for spatial model evaluation are then compared.
Requirements:
R programming is an advantage, interest in working with machine learning algorithms
References:
Brenning, A. (2012): Spatial cross-validation and bootstrap for the assessment of prediction rules in remote sensing: The R package sperrorest. IEEE International Geoscience and Remote Sensing Symposium.
Kühnlein, M., T. Appelhans, B. Thies, and T. Nauss, 2014: Precipitation Estimates from MSG SEVIRI Daytime, Nighttime, and Twilight Data with Random Forests. Journal of Applied Meteorology and Climatology, 53 (11), 2457–2480.
Meyer, H., Reudenbach, C., Hengl, T., Katurji, M., Nauss, T. (2018): Improving performance of spatio-temporal machine learning models using forward feature selection and target-oriented validation. Environmental Modelling & Software 101: 1-9.
Roberts, D. R., V. Bahn, S. Ciuti, M. S. Boyce, J. Elith, G. Guillera-Arroita, S. Hauenstein, J. J. Lahoz-Monfort, B. Schröder, W. Thuiller, D. I. Warton, B. A. Wintle, F. Hartig, and C. F. Dormann, 2017: Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography, doi:10.1111/ecog.02881.
Valavi R, Elith J, Lahoz‐Monfort JJ, Guillera‐Arroita G. blockCV: An r package for generating spatially or environmentally separated folds for k‐fold cross‐validation of species distribution models. Methods Ecol Evol. 2018;00:1–8. https://doi.org/10.1111/2041-210X.13107
Contact: Hanna Meyer
Author: Marc Dragunski
Supervisor: Hanna Meyer
Co-supervisor: Edzer Pebesma
SITCOMInference Attacks on Location Trajectories
Nowadays, smartphones are an omnipresent companion in our day-to-day life. With the ability to sense our location, location based services (LBS) have become widely used applications (e.g. for navigation, recommender systems, social networks, games, dating apps or fitness tracking). Hence, service providers collect vast amounts of location data about their users. Based on this collected data, providers or malicious third parties who gain access to this data, can infer a lot of additional information (e.g. home, workplace, shopping habits, religious beliefs, political views etc.) about somebody and thus harm their privacy. Those actions are called inferences or inference attacks.
The aim of this thesis is to research possible inference strategies based on the literature and to implement a selection of those. If the topic is chosen as a MSc thesis, an experimental evaluation of different inference strategies is also required.
Due to the algorithmic nature of this topic, the student should be interested in programming and not afraid of digging into some aspects of spatio-temporal analysis. The student is free to choose their programming language and environment of choice. Ideally, the software would be implemented in Javascript, so the results could be integrated into an existing learning application that is currently under development as part of the SIMPORT project (https://simport.net/). The aim of this learning application is to educate users about the risks and consequences of sharing their location data.
Contact: Sven Heitmann
Author: Eric Thieme-Garmann
Supervisor: Christian Kray
Co-supervisor: Sven Heitmann
- How do TensorFlow and PyTorch compare in terms of model fitting correspondences and the structure of intermediate layers when applied to EO data cubes?
- What are the differences in the interface and tooling availability between TensorFlow and PyTorch for spatio-temporal modeling, e.g. using ConvLSTM, TempCNN?
- How does the performance and interoperability of spatio-temporal deep learning models vary across different versions of TensorFlow and PyTorch when applied to EO data cubes?
NB : ONNX format can be used to port DL models between Tensorflow and Torch ; https://onnx.ai/
POC:
Dimension Denotation: https://onnx.ai/onnx/repo-docs/DimensionDenotation.html
Type Denotation: https://onnx.ai/onnx/repo-docs/TypeDenotation.html
Please contact:
Brian Pondi brian.pondi@uni-muenster.de
and
Edzer Pebesma edzer.pebesma@uni-muenster.de
Contact: Brian Pondi
Author: Ben Jannis Giese
Supervisor: Edzer Pebesma
Co-supervisor: Brian Pondi