Moritz Diener 

Background 

Foundation models — large AI models pre-trained on extensive, diverse data — are reshaping how we approach complex scientific and social challenges, including sustainability and the United Nations Sustainable Development Goals (SDGs). In environmental and geospatial domains, these models adapt to tasks such as satellite image segmentation, land cover classification, change detection, and environmental forecasting with far less labeled data than traditional approaches. Their ability to capture spatiotemporal patterns and cross-modal relationships makes them promising tools for filling critical data gaps in SDG monitoring, especially in regions where observations are scarce or costly to collect.

Because sustainability outcomes often depend on reliable interpretations of model outputs — for example in disaster response, biodiversity monitoring, or water resource management — integrating uncertainty estimation into foundation model pipelines is an important frontier. Quantifying uncertainty supports risk-aware decision-making and helps ensure that AI systems genuinely contribute to SDG progress rather than producing misleading or over-confident results under real-world variability.

Research Goal

The overarching goal of this thesis is to investigate how foundation models can be responsibly and effectively applied to sustainability-related challenges, with a particular focus but not limited to Earth observation and geospatial analysis.
The exact research questions, scope, and application domain will be defined jointly with the student, allowing the thesis to align with individual interests

Research Objectives

Depending on the student’s background and interests, the thesis may include one or more of the following objectives:

1. Foundation Models for Filling SDG Data Gaps: Explore how foundation models can be used to infer or enrich missing, sparse, or unevenly distributed data relevant to SDG indicators. Study transfer learning and generalization across regions, sensors, or time to support sustainability monitoring in low-resource settings.
2. Uncertainty Quantification for Foundation Model Outputs: Apply and develop uncertainty quantification methods for foundation model tasks, with a focus on dense prediction problems such as image segmentation. Analyze how uncertainty estimates relate to model errors, domain shifts, and data quality, and how they can support trustworthy decision-making in sustainability applications.
3. Benchmarking Novel Methods for Example Segmentation Tasks: Apply and evaluate emerging techniques such as visual prompting or prompt-based adaptation for segmentation in Earth observation imagery. 

Overall, the thesis is intended to be flexible and exploratory, encouraging students to shape a focused research question within this broader theme and to contribute both technically and conceptually to the responsible use of foundation models for sustainable development.

Your Profile

• Experience with Python 
• Interest in topics around sustainability and AI for good 
• Combining real-world challenges with science 

How to Apply

Please send your current transcript of records, a short CV, and a brief motivation (3–4 sentences) to: moritz.diener@kit.edu