Advancing automated georeferencing using gazetteers, geometry embeddings, and foundation models : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Computer Science, School of Mathematical and Computational Sciences, Massey University, Albany, Auckland, New Zealand

Abstract

Manual georeferencing is a highly labour intensive, time-consuming, and expensive process. As a result, vast amounts of valuable information remain locked up in archives in unstructured form. Biological collections are one prominent example: billions of un-georeferenced specimen records create major bottlenecks in mobilizing data for biodiversity research. The locations of such un-georeferenced specimen records are described only in text form, with descriptions like "Christchurch, near Halswell, Knights Stream, corner of Trices and Sabys Road, near Hohepa". Georeferencing this data is an essential step in studying the distribution of species. These challenges underscore the need for automated georeferencing methods.

Building automated georeferencing methods is inherently a multi-faceted task. Current implementations of such systems are mostly based on methods that can access existing geospatial knowledge bases (including gazetteers), and interpret aspects of complex textual spatial language and relations, but this thesis proposes georeferencing methods that can also comprehend maps, and reason over cartographic features as well as the textual spatial relation terms simultaneously. This research lays the ground work for such automation: investigating the current status of the gazetteers available for automated georeferencing, building innovative methods that enrich available geospatial knowledge bases and culminating in novel georeferencing methods and pipelines. The thesis, prepared within a "PhD-with-publication" format, is a part of the BioWhere project, a project that aims to georeference New Zealand's Biota from text. This is one of two PhD student theses under the BioWhere project.

Firstly, we investigate the current standing of digital gazetteers or place name knowledge bases, the foundational infrastructure on which many georeferencing system must rely. This survey synthesises decades of work on digital gazetteers, tracing their evolution from simple regional name-coordinate lists to complex geospatial knowledgebases that integrate authoritative datasets, volunteered geographic information (VGI), historical map extractions, and Linked Open Data. It also surveys methods for deduplication and record integration, highlighting challenges such as inconsistent feature coding, limited representation of place facets, heterogeneous geometry types, and fragmented approaches to data linking. We further discuss technology stacks used in gazetteer construction, including object relational databases, Linked Open Data, geospatial data formats, and semantic web technologies.

Secondly, we examine the problem of geospatial entity resolution, identifying database records that refer to the same real-world place, a critical operation in constructing unified place-name knowledge bases from heterogeneous sources. We introduce a novel framework, Omni, which addresses longstanding challenges in matching place records that differ across databases in naming conventions, feature-type coding schemes, and geometric representations. Unlike prior methods that reduce all geometries to point locations, thereby discarding essential spatial information, Omni incorporates an omni-geometry encoder capable of embedding points, lines, polylines, polygons, and multi-polygons into a unified representation space. We also present a manually annotated diverse-geometry dataset to assess the effectiveness of the omni-geometry encoder. The framework is evaluated across this dataset as well as several existing point-only datasets and demonstrates substantial improvements over existing pre-trained Language Model (PLM) based approaches. Omni produces up to 12% (F1) improvement over existing methods. The study also explores the capabilities of Large Language Models (LLMs) for this task, finding that while LLMs exhibit strong linguistic prior knowledge, their performance in diverse-geometry datasets trails that of Omni on all but one dataset partition.

Thirdly, we explore the potential of vision–language models (VLMs) to comprehend and reason with cartographic maps, introducing the map modality as a valuable channel for the purpose of georeferencing textual locality descriptions. In this work, we introduce a novel zero-shot method that augments textual prompts with labelled gridded map excerpts and relevant metadata about the grid, enabling the model to visually contextualize the spatial relation terms found in complex locality descriptions. To assess the effectiveness of this method, we manually create map excerpts for a small set of herbarium records from New Zealand. We further present a conceptual framework that can be used for the dynamic creation of such map excerpts. By combining locality descriptions with these gridded maps, the method tests whether a VLM can identify reference features, interpret relative spatial terms, and locate the most probable grid cell of a specimen’s collection site. Experiments show that this multimodal approach achieves substantially lower distance errors (~1 km average distance error) than text-only baselines and existing tools, revealing that VLMs can leverage both textual cues and visual spatial context even without task-specific training.

Finally, we develop a fully automated, end-to-end multimodal pipeline for georeferencing complex textual locality descriptions, integrating both text-only and multimodal foundation models, and structured geospatial knowledge bases into a unified framework. This work introduces a modular system that jointly performs spatial entity extraction, toponym refinement, toponym disambiguation, localized map generation, and visual–textual grounding. This framework automates the full workflow of producing gridded map excerpts and pairs them with locality descriptions to enable large-scale multimodal reasoning. We also produce a multi-modal dataset for cartographic map comprehension, generated directly through the pipeline, enabling fine-tuning of open-source VLMs. Experiments demonstrate that multimodal georeferencing substantially outperforms text-only and heuristic baselines, achieving up to threefold improvements in distance accuracy and locating over 90% of records within 10 km of ground truth when using a larger VLM backbone.

In summary, this thesis advances machine understanding of geographic space: how places, spatial relations, spatial language, and their cartographic context interact—and translates these advances into practical methods for automated georeferencing. These methods not only help unlock the vast spatial information stored in biological collections, improving biodiversity mapping and ecological research, but also provide a foundation for applying automated geospatial grounding to other text-rich domains such as historical archives, disaster reporting, and social media.