Building a foundation model for geospatial artificial intelligence

What is it about?

Foundation models (FMs) are large-scale pre-trained models that are task-agnostic; that is, they can be adapted to multiple downstream tasks via fine-tuning, few-shot, or even zero-shot learning. We explore the promises and challenges of developing multimodal FMs for geospatial artificial intelligence (GeoAI). We show that existing FMs (e.g. GPT-3) can outperform task-specific fully-supervised models on two geospatial semantics tasks with 2-9% improvement in a few-shot learning setting. Then, we show the limitations of these existing FMs on multiple GeoAI tasks, especially when dealing with geometries in conjunction with other modalities. Finally, we discuss the possibility of a multimodal FM for GeoAI that can reason over various types of geospatial data through geospatial alignments.

Featured Image

Photo by NASA on Unsplash

Photo by NASA on Unsplash

Why is it important?

(1) Despite the increasing popularity of foundation models (FMs) in the natural language and vision communities, we have yet to see their widespread adoption in GeoAI. Our vision paper highlights this very important research direction. (2) We are the first to show that with only 8 few-shot examples, such FMs (e.g., GPT-3) can outperform supervised, task-specific models on two geospatial semantics tasks with 2-9% performance improvements. (3) We highlight the challenges of developing FMs for GeoAI including the multimodal nature of GeoAI, geographic bias, diverse spatial scales, and geographic generalizability.

Perspectives