Embedded spatial systems
By 31 December 2025, a New Zealand geospatial practitioner could spend much of a working day dealing with geography without opening what an earlier generation would have recognised as a conventional GIS. Data could arrive through an API, sit in a spatial database, be processed by a scheduled workflow, appear in a browser application and be collected again through a mobile form. Satellite imagery could be classified with machine-learning methods, a point cloud could be processed into derived surfaces, and code for part of the workflow could be drafted with help from a general-purpose AI system. The underlying work was still geographic. It had simply spread into databases, web services, applications, software development and data engineering.
Organisations continued using GIS while adopting these additional tools. Geographic information became more useful as it escaped the specialist workstation and entered corporate databases, web maps, mobile devices, open-data services, cloud platforms and operational systems. Automation then moved repeated processing into scripts, SQL, ETL workflows and production pipelines. By the end of 2025, artificial intelligence was adding another layer to that environment, but it had not removed the older ones. A council could still need a desktop GIS, a national agency could still maintain a spatial database, and a field team could still need an offline map when the network stopped cooperating.
The difficult decisions remained with the practitioner. Someone still had to decide which dataset represented the thing being analysed, whether a coordinate system was appropriate, whether an address match was plausible, whether a classification was defensible and whether a result could safely be published. A model might identify a pattern faster than a person, yet poor training data remained poor training data and an unclear requirement remained unclear. A language model might produce a competent-looking Python script in seconds and still choose the wrong field or assume that every layer used the same projection. Competent-looking output was never the same thing as correct output. AI made that old GIS lesson easier to forget. The habit of checking what the computer had actually done remained useful.
By 2025, organisations were using a mixture of established systems and newer tools. New Zealand GIS in 2025 included long-lived operational databases, desktop software, open-source tools, cloud services, dashboards, remote sensing, mobile capture, automation and experimental AI methods at the same time. Different organisations adopted them at different speeds and for different reasons. Some machine-learning methods were already part of maintained national production. Some large-model work remained research. Some generative-AI use was appearing in practitioner tools and conference programmes without yet supplying a strong documentary basis for claiming organisation-wide geospatial adoption.
The same widening is visible in businesses that no longer fit neatly inside the traditional GIS label. Christchurch-based built a global business around Leapfrog 3D geological modelling and related subsurface data tools, then changed its name to in 2017 as it expanded beyond mining into civil, environmental and energy work. Bentley Systems acquired in 2021, by which time the New Zealand-born company was serving specialist users around the world. Digital spatial practice extended into adjacent fields using three-dimensional models, cloud data and location.
AI before the boom
New Zealand GIS researchers had used artificial-intelligence methods for decades before conversational AI became popular in the 2020s. At the first national multidisciplinary GIS conference in Wellington in June 1989, , and presented work integrating expert systems and GIS for land classification and assessment. A peer-reviewed follow-up appeared in the New Zealand Geographer in 1990. Those rule-based systems encoded geographic classification through computational reasoning before modern machine-learning methods.
Other lines followed. In 1990 presented a natural-language interface to a GIS database at the second annual Spatial Information Research Centre colloquium in Dunedin. By 1995 Sallis and were applying natural-language processing and GIS ideas to a railway safety-audit reporting system. In the 2000s New Zealand research included neural networks, fuzzy neural networks, support-vector methods and spatial data mining. 's 2003 doctoral research at used GIS and neural-network predictive modelling for river biological assessment, while 's SIRC work in 2005 and 2008 examined machine learning and rule extraction from spatial data.
The projects developed in separate research and operational settings. They came from different institutions, problems and technical traditions, and many were research rather than operational systems. New Zealand geospatial researchers had spent decades experimenting with ways to automate classification, prediction, language interpretation and decision support before large language models made artificial intelligence a daily topic in ordinary offices.
Learning from examples
Conventional GIS automation generally follows rules that a person has specified. A script can select features, calculate fields, transform coordinates and generate outputs because its instructions describe those operations. Machine learning changes part of that relationship. Instead of defining every rule explicitly, a practitioner supplies examples or other structured evidence from which a model learns patterns that can be applied to new data. The difference becomes practical when the task is easier to demonstrate repeatedly than to describe as a complete set of rules, as often happens in image classification.
New Zealand remote-sensing research provides a clear progression. In 2016 , and published work combining QuickBird imagery, LiDAR-derived topographic information and object-based classification to identify a native tree species, using methods including Random Forest and support-vector machines. In the Waikato, 's 2020 doctoral work used random forests, XGBoost and other methods with spatial yield and environmental data for precision agriculture. At Tauranga Harbour, and collaborators used Sentinel and Landsat imagery with ensemble machine learning, change detection and later radar-optical sensor fusion to map seagrass distribution and biomass. By 2021, researchers were using machine learning in geospatial work. It was one of several analytical approaches available for specific mapping problems.
The Land Use and Carbon Analysis System provides the stronger production case. In 2022 the Ministry for the Environment contracted to produce a 2020 national Land Use Map for greenhouse-gas reporting. The work used Sentinel-2 imagery and a mixture of established mapping methods and deep-learning techniques. The contract included reviewing areas mapped as forest in the 2016 Land Use Map using deep learning, identifying changes between mapping periods and incorporating those results into the maintained national time series. The final 2020 product was therefore not a laboratory demonstration detached from government use. It was part of an operational national reporting workflow.
The deep-learning work used a labelled training layer derived from the existing Land Use Map and Sentinel-2 imagery. Models were assessed through accuracy measures and visual inspection, predictions were reviewed for both model errors and errors in the existing map, and training could be refined and rerun. Final land-use changes were edited into the Ministry's geospatial feature service using ArcGIS Pro, with quality standards and manual review remaining part of the project. The computer helped identify and classify likely patterns at national scale. People still had to decide whether those patterns were acceptable evidence for a map used in international greenhouse-gas reporting.
The LUCAS workflow combined machine learning with established production methods. Training data carry earlier decisions about categories and boundaries. Satellite imagery varies with season, cloud, sensor and landscape. A model that performs well across one set of examples can perform poorly somewhere different. LUCAS worked because learned classification sat inside a larger production system that included existing datasets, controlled classes, imagery preparation, review, editing and quality assurance. The model formed part of a maintained geographic database.
Reading maps with AI
The other direction of change involved the maps themselves. A historical paper map can be scanned easily, but a scan is still an image until it is connected to geographic coordinates. Georeferencing has traditionally required a person to recognise features, identify corresponding control points and fit the image to a modern coordinate framework. That is manageable for a few maps and laborious for a large archive. In 2025, and at Victoria University of Wellington presented a different approach at GIScience 2025 in Christchurch.
Their paper, "Georeferencing Historical Maps at Scale", used salient line intersections to match features in historical map images with contemporary cadastral data. The method was designed for the awkward reality that old cadastral maps and current parcels do not contain identical geometry. Parcels have been subdivided, land use has changed, map images contain noise and historical maps may have uncertain orientation or scale. The researchers therefore looked for angular relationships in line intersections that persisted strongly enough to narrow possible locations. In one stage of evaluation, the algorithm was scaled across 100,000 randomly selected regions of Aotearoa New Zealand.
The research applied computational methods to historical maps and tested the resulting georeferencing against the source records. Pope and Frean explicitly frame historical cadastral maps within colonisation, dispossession and spatial data justice. Their method uses current LINZ parcel data to help locate older map images, but the old map and the modern cadastre are not treated as identical truths. The algorithm performs record linkage across two representations created at different times and for different purposes. Automation can make large collections more accessible while the provenance and politics of the source maps remain part of the interpretation.
By August 2025, New Zealand researchers were testing computational methods to search for geographic correspondences across national cadastral data and historical imagery at a scale impractical to inspect manually.
A model reads the map
The adjacent paper at GIScience 2025 pushed a different boundary. and of , with of Cardiff University, examined whether large multimodal models could combine a written locality description with a map. Natural-history collections contain huge numbers of specimens whose recorded locations are written as ordinary descriptions rather than coordinates. A collector might have written that a specimen came from a certain distance along a road, near a river, south of a settlement or beside a reserve. Converting such descriptions into coordinates has long required both language interpretation and geographic reasoning.
Their method presented a map excerpt divided into labelled grid cells alongside the locality description and asked a large multimodal model to identify the cell most likely to contain the described location. The work combined named-place extraction, spatial relations, map generation and multimodal reasoning. The researchers reported preliminary experiments on a small manually annotated dataset and an average distance error of roughly one kilometre for their approach, performing better than the comparison methods they tested. They also documented limitations, including difficulty following linear features and the tendency of vision models to confuse labels with mapped objects.
The complete workflow was still described as work in progress, although core elements were being used in experiments. By 2025, New Zealand GIScience researchers were testing a general multimodal model on tasks involving text and maps. The model generated coordinates and spatial inferences from language, beyond prose generation. It was being evaluated against a geographic answer.
The comparison with the 1990 SIRC natural-language work is instructive. Thirty-five years earlier, researchers were already interested in querying geographic information through ordinary language. The newer systems differed in scale, training and flexibility, but the underlying desire was familiar: reduce the distance between the way people describe place and the structured representations a GIS requires. Successive attempts to make computers work with geographic language have produced better methods, while facing the same difficulty: descriptions of place are often approximate, contextual and dependent on local knowledge.
Natural language returns
The 2010s add another bridge. A 2018 GIScience paper by , and examined the difficulties of creating annotated geospatial natural-language descriptions. In 2019 Stock and collaborators presented "He Tatai Whenua", work on automated extraction of landscape terms and meanings in New Zealand Māori. These projects predated the current large-language-model boom and treated language as data that needed careful annotation, interpretation and computational representation. Coordinates returned from natural-language queries required validation against the intended places.
The Māori language work addressed governance of language resources; the documented work did not establish a generative-AI deployment. Landscape terms can encode relationships and meanings that do not map neatly onto an English gazetteer or a generic ontology. Computational extraction can assist research, while authority over the source material and its future use still comes from people and communities rather than technical capability. By 2025, wider debates over Māori Data Sovereignty, cultural information and controlled access were already established in New Zealand geospatial practice. A model’s capacity to ingest text or maps left those obligations intact.
Generative systems changed the interaction again because a user could increasingly express a technical requirement in ordinary language and receive code, queries or explanations in return. For GIS practitioners this could mean a draft Python script, an SQL statement, a JavaScript function, an expression, a data-cleaning routine or help interpreting an error message. The resulting code might still perform entirely conventional deterministic processing. Asking a language model to write a buffer-and-join script is different from asking a multimodal model to infer a location from a map and a paragraph of text.
AI enters practice
By the end of 2025, generative AI was an emerging practitioner and research capability in geospatial work.
FOSS4G 2025, held in Auckland from 17 to 23 November, captures that boundary unusually well. The international open-source geospatial conference programme included workshops and presentations on geospatial AI applications, natural-language spatial search and AI-assisted programming. One session was explicitly titled "AI Coding and the Future of Open-Source Geospatial Software". Another dealt with semantic spatial search using natural language with PostGIS and vector search. These subjects appeared in the programme of a major geospatial conference held in New Zealand.
A local practitioner example appeared at FOSS4G Auckland in November 2025. presented a lightning talk about Delinify, a reporting tool he had recently built using open source and AI. The conference record documents a public demonstration; it does not describe organisational deployment or ongoing maintenance. The example places AI-assisted spatial reporting among the tools practitioners were exploring at the end of 2025.
AI-assisted coding sat in a similar position. By 2025 a practitioner could use a conversational coding assistant to produce candidate GIS code far faster than writing every line from memory. That lowered the practical barrier to scripting, particularly for people who understood their workflow but were not experienced software developers. It also moved the bottleneck. The difficult question became less "can I remember the syntax?" and more "does this code actually do what I intended?" A generated script that runs without an error can still join the wrong table, calculate in degrees instead of metres, ignore null values or quietly process only the first page of an API response.
AI-assisted coding extended earlier approaches to automating GIS work. The script, SQL query or application still becomes part of the same maintenance problem described in Chapter 43. It needs tests, documentation, controlled credentials, dependency management and someone who can diagnose it later. Generative systems made code easier to create, including code that a team might not previously have attempted. They did not make generated code self-validating or relieve an organisation of responsibility for what it ran.
Research and production
By the end of 2025, the phrase "using AI" covered too many different conditions to be historically useful on its own. LUCAS had deep-learning methods embedded within a completed national mapping project that supported formal greenhouse-gas reporting. The GIScience multimodal work was a peer-reviewed research experiment with a proposed workflow and clearly documented limitations. The historical-map georeferencing algorithm had been tested at large scale but remained research. The FOSS4G Auckland programme showed professional experimentation and product development. These are all legitimate parts of the history, but they are not equivalent stages of adoption.
The same caution applies to software capability. A commercial or open-source GIS package might expose a machine-learning tool, an AI service or a natural-language interface internationally. International availability alone provides no record of New Zealand use. An agency might issue a request for information about automatic feature extraction. A request for information records an investigation; a subsequent deployment requires separate documentation.
Earlier systems retain the terminology and technical context of their own period. Expert systems from 1989 belong to the AI lineage and differ fundamentally from large language models. Random Forest classification is machine learning, while generative AI is a different category. A Python script that extracts features according to explicit rules is automation. A language model that writes that Python script is an AI-assisted development tool, while the resulting workflow may remain deterministic.
The production boundary is also porous. Research methods can later become operational tools, and operational teams routinely borrow methods that began in universities. New Zealand's GIS history contains many examples of this movement, from university GIS teaching and SIRC research through remote sensing, web mapping and open-source development. The endpoint simply catches several AI-related methods at different positions along that path. Some were established enough to contribute to national production. Others were still being tested against the geography they were meant to understand.
AI governance
A highly automated spatial workflow can still be wrong about more than coordinates. It can use information that should not have been exposed, combine datasets under incompatible licences or produce a level of apparent precision that its sources do not support. Earlier chapters have dealt with privacy in health mapping, controlled operational information, Māori geographic knowledge and the distinction between open data and data that remains subject to authority or restriction. AI does not create a separate exemption from those rules. If anything, systems capable of ingesting large volumes of text, imagery and spatial data make provenance and permission harder to ignore.
The issue was visible in the same professional environment in which AI was becoming more prominent. FOSS4G Auckland 2025 included work on Indigenous data sovereignty and geoprivacy, including the MapSafe project and methods for protecting sensitive geographic information while still enabling analysis. MapSafe addresses geoprivacy rather than AI. Its relevance is the coexistence of two trends at the endpoint: tools were becoming more capable of combining and analysing information, while practitioners were also developing stronger technical methods for deciding what could safely be revealed.
This is especially important for Māori spatial information. The book has followed examples in which iwi and hapū use GIS to maintain cultural information, manage access and support their own decision-making. A generative or multimodal model may be technically capable of processing such data, but capability does not establish mandate. Training, indexing or exposing culturally sensitive information can create risks even where the source files were easy to obtain. The governing question remains who has authority over the information and for what purpose it may be used.
The same principle applies less dramatically across ordinary organisations. Sensitive infrastructure, personal addresses, environmental records and operational data can all acquire new exposure routes when they are copied into external systems or used in model workflows. A practitioner who once worried about sending a shapefile to the wrong email address now had additional questions about cloud services, model providers, prompts, retained inputs and generated outputs. The technology changed. The requirement to understand where the data went did not.
The specialist at the endpoint
The working definition of a GIS specialist had consequently become less tidy. Some roles remained close to cartography, spatial analysis, survey, remote sensing or database custodianship. Others crossed into Python, SQL, APIs, web development, data engineering, cloud platforms and automated deployment. A practitioner could spend most of a day writing transformations or configuring services and still be solving a recognisably geographic problem. Another could use sophisticated geospatial applications without writing code at all.
AI widened that range rather than collapsing it. A person with strong spatial knowledge could use a coding assistant to create scripts they might not have written unaided, but they still needed enough understanding to recognise when the script misunderstood the data. A developer could use a multimodal model in a geospatial application, but still needed to know how coordinates, map scale, feature geometry and spatial uncertainty affected the result. A remote-sensing specialist could train a classifier quickly, but still had to decide whether the training examples represented the landscapes to which the model would be applied.
The enduring skill lay in understanding geographic data, asking workable spatial questions, recognising uncertainty, structuring repeatable processes and knowing when the output deserved checking. New Zealand's GIS history had already moved through enough software generations to show the limits of memorising commands in one package. AI made some technical actions easier to initiate while those underlying judgements retained their value.
Spatial tools became part of other professions’ working systems. A planner might use a map application rather than a desktop GIS. An emergency manager might use a common operating picture. A scientist might call a spatial service from Python. A developer might query PostGIS through an API. A data analyst might ask an AI assistant to generate a spatial query. The user no longer had to identify the activity as GIS for spatial concepts and infrastructure to shape the result.
31 December 2025
At the end of 2025, New Zealand organisations used a range of platforms and working methods. The country had national spatial databases descended from earlier digitisation programmes, modern coordinate infrastructure, open data services, cloud-hosted applications, mobile capture, routinely available imagery, LiDAR and 3D data, enterprise GIS, automated pipelines and a growing body of machine-learning and multimodal research. Paper maps still existed. Desktop GIS still existed. So did specialised survey systems, spreadsheets, databases and all the awkward interfaces between them.
Organisations began using artificial intelligence alongside their existing GIS methods. Rule-based AI had appeared alongside GIS in the 1980s. Machine learning became useful for prediction and image classification over the following decades. Deep learning could assist a national land-use production workflow by the early 2020s. In 2025 New Zealand researchers were testing automatic historical-map georeferencing and large multimodal models that could combine maps with written locality descriptions, while practitioners were beginning to present AI-assisted spatial reporting and coding workflows in mainstream geospatial forums.
By 31 December 2025, research prototypes, coding assistants and multimodal map-reading models were still evolving. Research prototypes might become production services or disappear. Coding assistants might become ordinary infrastructure or be absorbed into other tools until the label became uninteresting. Models might improve at reading maps while continuing to fail in ways that require specialised checking. Practitioners had to evaluate those changing tools while maintaining established GIS systems.
Paper maps persisted alongside digital databases. Specialist systems persisted alongside desktop GIS. Desktop software persisted alongside web mapping. Commercial platforms persisted alongside open source. Offices and practitioners persisted through mobile GIS and automation. By the end of 2025 AI had begun changing how some geographic work could be classified, interpreted, programmed and delivered. The map remained one visible part of a much larger information system, and someone still had to know whether it was right.
Chapter source notes
1. The final chapter’s early AI lineage is supported by the documented 1989 GIS conference expert-system work, the 1990 SIRC natural-language GIS interface, subsequent New Zealand neural-network and spatial-data-mining research, and peer-reviewed remote-sensing studies. These examples establish a longer research lineage and are not presented as one continuous national AI programme.
2. Ministry for the Environment material on the LUCAS 2020 Land Use Map provides the strongest operational machine-learning case. Deep-learning techniques were used within a maintained national mapping workflow for greenhouse-gas reporting, alongside labelled training data, accuracy assessment, visual inspection, manual review and ArcGIS Pro editing. The chapter therefore avoids the misleading shorthand that “AI mapped New Zealand”.
3. Rere-No-A-Rangi Pope and Marcus Frean, “Georeferencing Historical Maps at Scale”, GIScience 2025, is the principal automatic historical-map georeferencing case. It uses salient line intersections and contemporary cadastral data and was evaluated across large samples of Aotearoa New Zealand. It is peer-reviewed research, not an operational national georeferencing service.
4. Kalana Wijegunarathna, Kristin Stock and Christopher B. Jones, GIScience 2025, supports the large-multimodal-model locality-georeferencing experiment combining written descriptions with map excerpts. The paper reports preliminary experiments on a small manually annotated dataset and describes the complete workflow as work in progress. It is evidence of research capability at the cutoff, not institutional production deployment.
5. The official FOSS4G Auckland 2025 programme supports the presence of geospatial AI, natural-language search and AI-assisted programming in mainstream professional discussion by November 2025. Phil Clunies-Ross’s lightning talk on AI-assisted spatial reporting and Delinify provides a narrower New Zealand practitioner signal. The evidence establishes practitioner product development and experimentation, not organisation-wide generative-AI adoption.
6. The sources separately document research, maintained production systems and practitioner experiments. Expert systems are not large language models, Random Forest classification is machine learning but not generative AI, deterministic scripts remain automation even when an AI system helped draft the code, and software availability is not evidence of New Zealand deployment.