Poly2Vec: Polymorphic Fourier-Based Encoding of Geospatial Objects for GeoAI Applications

Area of Science:

• Geospatial artificial intelligence (GeoAI)
• Machine learning (ML)
• Geospatial data representation

Background:

• Encoding geospatial objects is crucial for GeoAI, but common methods lose vital spatial information.
• Existing object-specific encoders lack versatility for diverse geospatial data types (points, polylines, polygons).

Purpose of the Study:

• To introduce Poly2Vec, a unified, polymorphic encoding approach for geospatial objects.
• To preserve essential spatial properties (topology, direction, distance) during encoding.
• To enhance GeoAI workflows with a versatile and effective geospatial object representation.

Main Methods:

• Developed Poly2Vec, a Fourier-based encoding method for unifying geospatial objects.
• Incorporated a learned fusion module to adaptively integrate Fourier transform magnitude and phase.
• Evaluated Poly2Vec on five diverse tasks, including spatial relationship preservation and GeoAI workflow integration.

Main Results:

• Poly2Vec outperformed object-specific baselines in preserving topology, direction, and distance.
• Integration of Poly2Vec improved performance in population prediction and land use inference tasks.
• Demonstrated consistent superiority across diverse geospatial tasks and data types.

Conclusions:

• Poly2Vec offers a robust and unified solution for geospatial object encoding in GeoAI.
• The method effectively preserves critical spatial information lost in traditional approaches.
• Poly2Vec enhances the performance and applicability of machine learning in geospatial analysis.