SpatialVec: A Unified Representation Learning Framework for Geospatial Objects with Boundary-Aware Sampling

Samya Acharja; Kanchan Chowdhury

cs.CVOct 28, 2026

SpatialVec: A Unified Representation Learning Framework for Geospatial Objects with Boundary-Aware Sampling

Authors: Samya Acharja, Kanchan Chowdhury

Organizations: Marquette University, Milwaukee, USA

Abstract

Spatial representation learning is essential for Geospatial Artificial Intelligence (GeoAI) applications, enabling machine learning models to capture the geometry, location, and spatial relationships of geographic entities such as points, polylines, and polygons. Existing spatial encoding methods often struggle to preserve fine-grained topological information, particularly for interactions involving complex geometries such as polylines and polygons. These approaches inadequately capture boundary structure and local spatial context, causing topology-sensitive relationships near geometry boundaries to be underrepresented in the learned embedding space. Consequently, accurately modeling complex spatial interactions remains a significant challenge for downstream spatial reasoning tasks. To fill the gap, we introduce SpatialVec, a boundary-aware spatial representation learning framework that adaptively samples geometry boundaries and their surrounding regions, capturing fine-grained geometric structures critical for topology-aware spatial reasoning. A transformer-based encoder-decoder trained through geometry reconstruction produces compact, geometry-aware, and unified representations for points, polylines, and polygons. Empirical evaluation on two real-world datasets shows that SpatialVec substantially improves the preservation of challenging topological relationships involving polygonal interactions and outperforms state-of-the-art baselines significantly on downstream prediction and classification tasks.

Figures & tables

Fig. 1: End-to-end data flow and architecture of SpatialVec framework.

	Singapore					NYC
Method	Pt–Pl $\uparrow$	Pt–Pg $\uparrow$	Pl–Pl $\uparrow$	Pl–Pg $\uparrow$	Pg–Pg $\uparrow$	Pt–Pl $\uparrow$	Pt–Pg $\uparrow$	Pl–Pl $\uparrow$	Pl–Pg $\uparrow$	Pg–Pg $\uparrow$
DIRECT	82	84	73	36	35	84	90	74	49	44
TILE	79	70	50	45	41	65	78	50	49	40
WARP	88	88	71	47	47	88	88	73	55	38
GRID	84	84	69	45	45	82	89	73	51	38
THEORY	89	90	71	45	45	89	90	73	59	45

TABLE I: Model accuracy on topological relationship classification. All values are scaled by

\times 10^{-2}

	Singapore			NYC
Method	Pg–Pt $\downarrow$	Pl–Pg $\downarrow$	Pg–Pg $\downarrow$	Pg–Pt $\downarrow$	Pl–Pg $\downarrow$	Pg–Pg $\downarrow$
TILE	99.9	115.5	114.3	127.6	154.3	167.4
THEORY	24.30	25.00	25.00	26.20	26.60	27.40
Hist2Vec	16.20	22.47	30.31	21.80	27.42	32.44
Poly2Vec	13.10	21.00	15.00	28.07	28.05	52.70
Geo2Vec	6.40	13.00	5.40	10.20	13.00	12.09

TABLE II: Comparing MAE on distance estimation. All MAE values are scaled by

\times 10^{-3}

Method	NYC $\downarrow$	Singapore $\downarrow$
Hist2Vec	0.4448	0.2931
Poly2Vec	0.4430	0.3130
Geo2Vec	0.4562	0.3523
SpatialVec	0.3919	0.2765

TABLE III: Comparing WMAE on range-query’s selectivity prediction.

Method	Preprocessing Time	Training Time	Total Time
Hist2Vec	0.3 min	250.00 min	250.3 min
Poly2Vec	0.1 min	300 min	300.01 min
Geo2Vec	6.4 min	473.4 min	479.8 min
SpatialVec	2.4 min	507.6 min	510.00 min

TABLE IV: Comparing embedding preprocessing and training latency.

Panel of Figure 6 — Fig. 2: Embedding-generation latency with increasing polygon complexity on the NYC dataset. The y-axis uses a logarithmic scale.

Variant	Pt–Pl	Pt–Pg	Pl–Pl	Pl–Pg	Pg–Pg
Full SpatialVec	99	99	98	80	86
w/o boundary-aware sampling	97	98	95	79	75
w/o occupancy and boundary fields	97	98	98	72	75
w/o metadata and location integration	98	98	97	78	76
w/o boundary-aware pooling	95	97	96	74	75
w/o multi-view aggregation	96	98	94	73	76

TABLE V: Accuracy on topological relationship classification on NYC dataset for SpatialVec variants.

Geospatial Representation Learning

Geospatial representation learning builds reusable embeddings of places, imagery, maps, mobility, and Earth data for spatial prediction and reasoning.

See all 137 papers

May 12, 2026

NARA: Anchor-Conditioned Relation-Aware Contextualization of Heterogeneous Geoentities

Geospatial foundation models have primarily focused on raster data such as satellite imagery, where self-supervised learning has been widely studied. Vector geospatial data instead represent the world as discrete geoentities with explicit geometry, semantics, and structured spatial relations, including metric proximity and topological relationships. These relations jointly determine how entities interact within space, yet existing representation learning methods remain fragmented, often restricted to specific geometry types or partial spatial relations, limiting their ability to capture unified spatial context across heterogeneous geoentities. We propose NARA (Neural Anchor-conditioned Relation-Aware representation learning), a self-supervised framework for vector geoentities. NARA learns context-dependent representations by jointly modeling semantics, geometry, and spatial relations within a unified framework and captures relational spatial structure beyond proximity alone, enabling rich contextualized representations across heterogeneous geoentities of points, polylines, and polygons. Evaluation on building function classification, traffic speed prediction, and next point-of-interest recommendation shows consistent improvements over prior methods, highlighting the benefit of unified relational modeling for vector geospatial data.

Jina Kim, Gengchen Mai, Lingyi Zhao +2

Department of Computer Science and Engineering, University of Minnesota, Minnesota, USA · Department of Geography and the Environment, University of Texas at Austin, Texas, USA · Novateur Research Solutions, Virginia, USA

Geospatial Representation Learning Unsupervised Representation Learning

Ask PDF

Oct 4, 2026

LEON: Location Embeddings from OSM Neighborhoods via Hexagonal Graph Masked Autoencoders

Geographic information systems increasingly rely on sophisticated spatial representation learning techniques to extract meaningful patterns from complex geospatial data. This paper introduces LEON, a novel self-supervised framework that adapts Graph Masked Autoencoders (GraphMAE) for geospatial region representation learning. Our method leverages the inherent spatial structure of geographic data by constructing hexagonal grid graphs using H3 indexing and applying masked autoencoding techniques to learn robust spatial embeddings from OpenStreetMap (OSM) amenity distribution patterns. We evaluate LEON on multiple real-world datasets including EuroSAT satellite imagery classification and various geographic prediction tasks (housing prices, crime prediction, and urban analytics). Experimental results demonstrate that LEON achieves significant improvements in spatial understanding, with up to 1.87% accuracy improvement on EuroSAT classification and consistent performance gains across geographic prediction benchmarks. The learned embeddings exhibit highly structured and distinct properties, making them particularly suitable for downstream spatial analysis tasks. Our findings suggest that self-supervised learning provides an effective paradigm for geospatial region representation learning using widely available crowdsourced data.

Szymon Soltysiak, Radoslaw Malek, Jedrzej Kusnierz +3

Department of Artificial Intelligence, Wroclaw University of Science and Technology, Wroclaw, Poland

Masked Autoencoders Geospatial Representation Learning Graph SSL

Ask PDF

Oct 12, 2026

Geospatial Vector Geometry as a Native Spatial Modality for Vision-Language Models

Geospatial reasoning links VLMs to real-world action and decision making. Reliable reasoning requires accurate spatial information and effective spatial representations. However, current VLMs obtain this information mainly from language and vision. At larger geographic scales, limited visual observations make it difficult to establish accurate spatial references and recover precise spatial relations. This leads to a bottleneck for VLM spatial perception and reasoning. This paper proposes that geospatial vector geometry can become a native spatial modality within a VLM and introduces GeoVec-VLM. First, we develop a structured geometry representation that preserves the spatial structure and arrangement of vector objects, building an explicit spatial basis. Second, we introduce a geometry interaction strategy that brings this basis into multimodal computation, where it interacts with visual content and language references according to the queried spatial relations. Together, these designs allow perception and reasoning to operate on shared geometric references. Across five spatial reasoning tasks, GeoVec-VLM improves mean accuracy by about 11% relative to the strongest baseline, maintains its gains under geographic transfer, and shows greater geometric consistency. Our results show that geospatial vector geometry offers a promising internal spatial basis for VLM spatial reasoning. https://github.com/TXFR/GeoVector-VLM

Huang Fei, Boyi Chen, Hao Xue +3

The Hong Kong University of Science and Technology (Guangzhou), China · The Hong Kong University of Science and Technology, Hong Kong SAR, China

Vision-Language Models Geospatial Reasoning Geospatial Representation Learning Vision-Language Grounding

Ask PDF

Abstract

Geospatial Representation Learning

Explore similar work

NARA: Anchor-Conditioned Relation-Aware Contextualization of Heterogeneous Geoentities

LEON: Location Embeddings from OSM Neighborhoods via Hexagonal Graph Masked Autoencoders

Geospatial Vector Geometry as a Native Spatial Modality for Vision-Language Models