Spatially Explicit Artificial Intelligence Lab

Our Mission - As Michael Goodchild puts it, “A model is said to be spatially explicit when it differentiates behaviors and predictions according to spatial locations”. In the context of machine learning (ML) and artificial intelligence (AI), we see the need for spatially explicit models focusing on better ways to design ML/AI model architectures for consuming spatial information by incorporating spatial knowledge and spatial inductive bias such as spatial heterogeneity, spatial autocorrelation, map projection, and so on. The recent decade has witnessed many advancements of spatially explicit ML and AI technology for various GeoAI problems such as geographic question answering, geographic knowledge graph summarization, POI distribution modeling, trajectory prediction, traffic forecasting, geo-aware fine-grained species recognition, building pattern recognition, map generalization, remote sensing image classification, terrain feature recognition, and so on. However, adding spatial inductive bias to make spatially explicit models will yield more complex ML/AI model architectures. Investigating the trade-off between designing a spatially explicit ML/AI architecture versus a more general setup that would have to learn to value space implicitly is an important research question for the spatial data science and GeoAI community. The recent advancement of various foundation models (e.g., ChatGPT, GPT-4 Vision) also huge us think about what additional benefits spatial thinking can bring to the table if a large foundation model can already achieve superior zero-shot or few-shot performance on various geospatial tasks. The mission of SEAI lab is to develop a set of spatially explicit artificial intelligence models, tools, as well as their associated datasets to demonstrate the advantage of spatial thinking in novel GeoAI model design and guide the development of the entire GeoAI and Spatial Data Science domain.

Our Projects