From R-Trees to Learned Indexes: A Systematic Review of Spatial Indexing Across 2D, 3D and Point-Cloud Geospatial Query Scenarios

Abstract

Learned indexes are presented as a new direction for spatial data access, yet their transfer to native 3D geospatial and point-cloud tasks remains unclear. Building on existing reviews and benchmarks, this review connects index structures with geospatial application scenarios, query requirements, and benchmark dimensions. Following PRISMA 2020 reporting guidance, it combines searches of Scopus, Web of Science, IEEE Xplore and ACM Digital Library with revision-stage reference checking. After eligibility assessment and consolidation of overlapping reports, all 111 retained studies were included in the qualitative synthesis. The review develops a unified taxonomy of traditional and learned spatial indexes, including hybrid designs, along two analytical dimensions, rule-driven and learning-driven designs, and 2D and 3D data support. The synthesis reveals an imbalance: learned indexes are widely evaluated in 2D and general multidimensional settings, whereas native 3D geospatial indexing remains dominated by traditional and hybrid structures. The review reframes the learned–3D gap as a workload-adaptation and benchmark-standardization problem. Finally, it identifies design requirements and benchmark dimensions for geospatial-aware learned 3D indexes, including density-aware partitioning, geometry-preserving learning, exact–approximate refinement, update-aware maintenance, out-of-core scalability, and system integration.