ISPRS International Journal of Geo-Information

Geospatial data contains abundant temporal, spatial, and attribute information related to geographic entities, serving as a critical foundation for the development of new quality productive forces. Achieving effective semantic representation of geographic features is a key prerequisite for enabling intelligent data analysis and application. Spatial relationships are an essential semantic characteristic of geographic features and are among the most commonly introduced characteristics for representation learning in recent studies. However, current research primarily focuses on point-based features (such as POIs) or on converting line and polygon features into point features, leading to inaccuracies in representing spatial proximity between features, which in turn affects the representation outcomes. In this paper, a feature unit delineation method based on geometric shapes is proposed to better construct the proximity relationship graph between features. It enriches the semantic features of feature units using a large language model and trains an edge optimization model to refine the graph structure. Building upon iGCL, a semantic representation learning model for geographic features is constructed. Finally, using Xiamen Island and Nanchang City as study areas, experimental datasets are built by integrating OSM and Amap (Gaode) POI data. The performance of the proposed model is validated and compared with other models through three tasks: semantic overlap calculation, node classification, and node clustering. The results show that the feature representations obtained by the proposed model achieve the best performance across all tasks, fully demonstrating its effectiveness and superiority in geographic feature representation learning, thereby providing a foundation for subsequent intelligent geographic feature tasks.

Location-based services (LBS) have traditionally helped users answer questions such as where to go or what is nearby. With the emergence of large language model (LLM)-based agents, LBS are beginning to move toward understanding open-ended goals, coordinating spatial tasks, and supporting decisions about what to do and how to do it. This review refers to LBS that integrate such agentic capabilities into location-dependent service processes as Agentic LBS. To systematically characterize the current state of Agentic LBS and identify directions for future research, we conducted a scoping review and proposed an Autonomy–Innovation–Experience (A-I-E) framework as its analytical lens. Following the PRISMA Extension for Scoping Reviews (PRISMA-ScR), we searched the ACM Digital Library, Web of Science, and Scopus for peer-reviewed, English-language journal articles and conference papers published between 2021 and 2026. The search identified 228 records. After removing 21 duplicates, 207 records underwent screening, and 27 studies were included in the final review. The included studies were systematically charted and analyzed through descriptive mapping and qualitative coding to characterize their application directions, autonomy levels, innovation patterns, and user-experience implications. The results show that: (1) Agentic LBS research remains concentrated at lower autonomy levels, with 23 of the 27 studies (85.18%) classified as L1–L2 and four as L3; nevertheless, the emergence of L3 systems indicates a gradual movement toward more complex task planning and coordination. (2) LLM-based agents are expanding LBS beyond location-based information retrieval toward more personalized, context-aware, and task-oriented service delivery. (3) Reported user-experience benefits are concentrated in usability, accessibility, personalization, and task efficiency, whereas evidence concerning trust, user control, privacy, safety, and long-term use remains limited. (4) By synthesizing these findings across autonomy, innovation, and experience, this review proposes a future-oriented conceptual framework for understanding the emerging trajectory of Agentic LBS.

Driving simulation needs labeled 3D scenes, but mobile mapping system (MMS) light detection and ranging (LiDAR) is often delivered unclassified, without point-wise labels for training. We present a training-free pipeline that transfers the semantics of a vector high-definition (HD) map, the Korean precision road map, onto MMS LiDAR and outputs a labeled 0.5 m voxel scene, and we analyze what limits its accuracy. Each voxel takes the label of the nearest map feature whose category-specific radius and elevation window it satisfies, with elevations measured from a ground model built from the map’s 3D vertices. The pipeline needs no training data but requires a 3D vector map. On an expressway interchange, the ground model reached a median error of 6.6 cm, on par with inverse-distance weighting (IDW), triangulated irregular network (TIN), and local-plane interpolation. A blind, stratified random sample of 695 cells and a blind review of all 91 output objects showed that 98.0% [95.8, 99.5] of the output voxels are static solid structure and 73% [63, 81] of the objects are the mapped facility; the cell-level macro-F1 over six categories was 0.44 [0.39, 0.52] (guardrail 0.73, tunnel 0.94, sign 0.14). Errors were concentrated: four tunnel-mounted signals labeled the tunnel ceiling, and the radii of signs and posts captured roadside vegetation. The results show where map-based labels are reliable for simulation and which map content needs treatment.