SPOK: tokenizing geographic space for enhanced spatial reasoning in GeoAI

Abstract

Tokenization is a fundamental process that feeds data into artificial intelligence (AI) models by breaking data into discrete, learnable units known as tokens. For instance, tokenizing text supports GPT in understanding semantic and contextual relationships within language. However, geospatial tokenization, the process of transforming spatial-temporal data into discrete tokens for geographic modeling, remains significantly underexplored, which has constrained the spatial reasoning capabilities of AI models. This article presents SPOK, a geospatial tokenization approach that explicitly encodes spatial relationships into spatial tokens. Using urban space as an example, SPOK defines parcels delimitated by road links as the basic spatial tokens and employs dynamic location referencing to encode their relative spatial relationships. By embedding these spatial relationships into high-dimensional vectors, SPOK effectively queries and models interactions among spatial tokens in latent space, thereby enhancing spatial reasoning in urban environments. We evaluate SPOK through dynamic mobility flow prediction and the results demonstrate that SPOK can infer origin-destination (OD) patterns without prior spatial interaction data, outperforming baselines with up to 20% reduction in RMSE and 14% increase in (Formula presented.) By offering how spatial relationships can be tokenized, this study would like to call for attention on geospatial tokenization to develop geospatial foundation models and GeoAI.