Adaptive graph construction and graph contrastive learning for geospatial feature embedding

Geospatial features serve as a cornerstone in urban planning and location-based services. Effective geospatial feature embedding is crucial for downstream geospatial Deep Neural Networks (DNN) applications. Currently, geospatial feature representation learning approaches rely on image or textual data to implicitly encode specific geospatial features, such as Points of Interest (POIs), roads, and urban areas, and struggle to effectively model spatial proximity relationships between imbalanced-distribution geospatial features. Moreover, the difficulty in effectively selecting appropriate geographic feature negative samples during model training poses a challenge to traditional representation learning methods. To address these limitations, we propose a novel Geospatial Adaptive Contrastive Embedding approach to extract neighboring features in an imbalanced-distribution space using Delaunay triangulation. It can effectively capture the natural proximity relationship through analysis the distribution characteristics of the geospatial features without setting a threshold. Graph contrastive learning is employed to encode geospatial features into low-dimensional feature vectors through self-supervised spatial similarity preservation. Finally, a comprehensive evaluation of the generated spatial proximity graphs and the learned representation are performed by using through three real-world geospatial datasets. The results demonstrate that our approach successfully learns an invariant and discriminative representation of geospatial features by constructing an effective spatial proximity graph.