About me
Bonjour! I’m Xiayin Lou, a doctoral student in Chair of Cartography and Visual Analytics at Technical University of Munich (TUM). GeoAI shows powerful ability in solving geospatial problems, but drawbacks still remain. As a result, knowing how much we can trust GeoAI is important. I study the bias in GeoAI from a statistical perspective, which consists four connected topics:
- Uncertainty/bias quantification for GeoAI models: Designing methodology that identifies whether GeoAI predictions should (not) trusted and whether GeoAI predictions perform differently across different locations. Recent focus is more about using conformal inference in the geospatial context. Applications include urban planning, epidemiology, responsible use of LLMs, etc.
- Generalization and mechanism of GeoAI models: Studying the generalization and its underlying mechanism of GeoAI models across datasets, population, and regions. Related themes concern distribution shift among different geographical regions.
- Geosaptial visual analytics for GeoAI models: Understanding GeoAI models through interactive geospatial visual analytics. Related themes concern human-in-the-loop exploration of model behavior, spatial prediction patterns, and uncertainty in GeoAI systems.
- LLM-powered geospatial knowledge discovery: Developing a self-evolving framework for GeoAI that can automatically optimize a geospatial algorithms. This can large accelerate the geospatial scientific research.
During my leisure time, I usually enjoy trail running, marathon, rowing, etc.
News
- Dec 2025: How can we build more reliable and trustworthy spatial intelligence? Happy to share our short paper “Towards the Uncertainty-aware Geospatial Artificial Intelligence,” exploring the future of uncertainty in GeoAI!
- Oct 2025: Can we truly trust the explanations provided by AI for spatial data? Our new paper “GeoXCP: uncertainty quantification of spatial explanations in explainable AI“ is now available online in International Journal of Geographical Information Science, providing a framework to quantify XAI uncertainty.
- Sep 2025: Can AI autonomously “think” like a geographer to discover new spatial algorithms? Excited to share our preprint “GeoEvolve: Automating Geospatial Model Discovery via Multi-Agent Large Language Models,” a framework that couples evolutionary search with domain knowledge to automatically refine classical geospatial models like Kriging and GWR!
- Jul 2025: Looking for a model-agnostic way to measure the reliability of spatial predictions? We are excited to announce that “GeoConformal prediction: a model-agnostic framework of measuring the uncertainty of spatial prediction“ is now published in Annals of the American Association of Geographers.