As geospatial data from web platforms becomes increasingly accessible and regularly updated, urban representation learning has emerged as a critical research area for advancing urban planning. Recent studies have developed foundation model-based algorithms to leverage this data for various urban-related downstream tasks. However, current research has inadequately explored deep integration strategies for multiscale, multimodal urban data in the context of urban foundation models. This gap arises primarily because the relationships between micro-scale (e.g., individual points of interest and street view imagery) and macro-scale (e.g., region-wide satellite imagery) urban features are inherently implicit and highly complex, making traditional interaction modeling insufficient. This paper introduces a novel research problem – how to learn multiscale urban representations by integrating diverse geographic data modalities and modeling complex multimodal relationships across different spatial scales. To address this significant challenge, we propose UrbanMFM, a spatial graph-based multiscale foundation model framework explicitly designed to capture and leverage these intricate relationships. UrbanMFM utilizes a self-supervised learning paradigm that integrates diverse geographic data modalities, including POI data and urban imagery, through novel contrastive learning objectives and advanced sampling techniques. By explicitly modeling spatial graphs to represent complex multiscale urban relationships, UrbanMFM effectively facilitates deep interactions between multimodal data sources. Extensive experiments on datasets from Singapore, New York, and Beijing demonstrate that UrbanMFM outperforms the strongest baselines significantly in four representative downstream tasks. By effectively modelling spatial hierarchies with diverse data, UrbanMFM provides a more comprehensive and adaptable representation of urban environments.

Bimodal data, such as image-text pairs, has become increasingly prevalent in the digital era. The Hybrid Vector Query (HVQ) is an effective approach for querying such data and has recently garnered considerable attention from researchers. It calculates similarity scores for objects represented by two vectors using a weighted sum of each individual vector's similarity, with a query-specific parameter α to determine the weight. Existing methods for HVQ typically construct Approximate Nearest Neighbors Search (ANNS) indexes with a fixed α value. This leads to significant performance degradation when the query's α dynamically changes based on the different scenarios and needs. In this study, we introduce the Dynamic Edge Navigation Graph ( DEG ), a graph-based ANNS index that maintains efficiency and accuracy with changing α values. It includes three novel components: (1) a greedy Pareto frontier search algorithm to compute a candidate neighbor set for each node, which comprises the node's approximate nearest neighbors for all possible α values; (2) a dynamic edge pruning strategy to determine the final edges from the candidate set and assign each edge an active range. This active range enables the dynamic use of the Relative Neighborhood Graph's pruning strategy based on the query's α values, skipping redundant edges at query time and achieving a better accuracy-efficiency trade-off; and (3) an edge seed method that accelerates the querying process. Extensive experiments on real-world datasets show that DEG demonstrates superior performance compared to existing methods under varying α values.

Pre-trained Foundation Models (PFMs) have ushered in a paradigm-shift in AI, due to their ability to learn general-purpose representations that can be readily employed in downstream tasks. While PFMs have been successfully adopted in various fields such as NLP and Computer Vision, their capacity in handling geospatial data remains limited. This can be attributed to the intrinsic heterogeneity of such data, which encompasses different types, including points, segments and regions, as well as multiple information modalities. The proliferation of Volunteered Geographic Information initiatives, like OpenStreetMap, unveils a promising opportunity to bridge this gap. In this paper, we present CityFM, a self-supervised framework to train a foundation model within a selected geographical area. CityFM relies solely on open data from OSM, and produces multimodal representations, incorporating spatial, visual, and textual information. We analyse the entity representations generated by our foundation models from a qualitative perspective, and conduct experiments on road, building, and region-level downstream tasks. In all the experiments, CityFM achieves performance superior to, or on par with, application-specific algorithms.

The task of address matching involves linking unstructured addresses to standard ones in a database. The challenges presented by this task are manifold: misspellings, incomplete information, and variations in address content are some examples. While there have been previous studies on entity matching in natural language processing, for the address matching solution, existing approaches still rely on string-based similarity matching or manually-designed rules. In this paper, we propose StructAM, a novel method based on pre-trained language models (LMs) and graph neural networks to extract the textual and structured information of the addresses. The proposed method leverages the knowledge acquired by large language models during the pre-training phase, and refines it during the fine-tuning process on the address domain, to obtain address-specific semantic features. Meanwhile, it also applies an attribute attention mechanism based on Graph Sampling and Aggregation (GraphSAGE) module to capture internal hierarchy information of the address text. To further enhance the accuracy of our algorithm in dirty settings, we incorporate spatial coordinates and contextual information from the surrounding area as auxiliary guidance. We conduct extensive experiments on real-world datasets from four different countries and the results show that StructAM outperforms state-of-the-art baseline approaches for address matching.

A geospatial Knowledge Graph (KG) is a heterogeneous information network, capable of representing relationships between spatial entities in a machine-interpretable format, and has tremendous applications in logistics and social networks. Existing efforts to build a geospatial KG, have mainly used sparse spatial relationships, e.g., a district located inside a city, which provide only marginal benefits compared to a traditional database. In spite of the substantial advances in the tasks of link prediction and knowledge graph completion, identifying geospatial relationships remains challenging, particularly due to the fact that spatial entities are represented with single-point geometries, and textual attributes are frequently missing. In this study, we present GTMiner, a novel framework capable of jointly modeling Geospatial and Textual information to construct a knowledge graph, by mining three useful spatial relationships from a geospatial database, in an end-to-end fashion. The system is divided into three components: (1) a Candidate Selection module, to efficiently select a small number of candidate pairs; (2) a Relation Prediction component to predict spatial relationships between the entities; (3) a KG Refinement procedure, to improve both coverage and correctness of a geospatial knowledge graph. We carry out experiments on four cities' geospatial databases, from publicly-available sources and compare with existing algorithms for link prediction and geospatial data integration. Finally, we conduct an ablation study to motivate our design choices and an efficiency analysis to show that the time required by GTMiner for training and inference is comparable, or even shorter, than existing solutions.

A geospatial database is today at the core of an ever increasing number of services. Building and maintaining it remains challenging due to the need to merge information from multiple providers. Entity Resolution (ER) consists of finding entity mentions from different sources that refer to the same real world entity. In geospatial ER, entities are often represented using different schemes and are subject to incomplete information and inaccurate location, making ER and deduplication daunting tasks. While tremendous advances have been made in traditional entity resolution and natural language processing, geospatial data integration approaches still heavily rely on static similarity measures and human-designed rules. In order to achieve automatic linking of geospatial data, a unified representation of entities with heterogeneous attributes and their geographical context, is needed. To this end, we propose Geo-ER1, a joint framework that combines Transformer-based language models, that have been successfully applied in ER, with a novel learning-based architecture to represent the geospatial character of the entity. Different from existing solutions, Geo-ER does not rely on pre-defined rules and is able to capture information from surrounding entities in order to make context-based, accurate predictions. Extensive experiments on eight real world datasets demonstrate the effectiveness of our solution over state-of-the-art methods. Moreover, Geo-ER proves to be robust in settings where there is no available training data for a specific city.