BRENNEN

Industries / Technology / Mapping and geospatial

GIS software

GIS software (geographic information systems) stores map layers as data so they can be overlaid, measured and analyzed, and serves them to screens and the web; the work covers desktop programs, spatial databases, shared code libraries, and the file formats and standards that let them exchange data.

7

Milestones achieved, from 1965 to 2023.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1965

    A lab for computer maps

    After a short course on computer mapping in 1963, the architect Howard Fisher, working with the programmer Betty Benson, wrote SYMAP at Northwestern University in 1964, one of the first programs for drawing maps by computer. In 1965, with a grant from the Ford Foundation, he set up the Laboratory for Computer Graphics at Harvard, where Esri says many early ideas of GIS were worked out. Jack Dangermond, a member of the lab, founded Esri with his wife Laura in 1969.

    esri.comesri.in

    Credit

    Howard Fisher; Betty Benson; the Ford Foundation

  2. 1981

    GIS sold as a product

    Esri began in 1969 as a consulting firm that applied computer mapping to land-use planning for clients. It turned the tools built on those projects into ARC/INFO, released in 1981, which Esri calls the first commercial GIS product, and the release turned Esri into a software company.

    esri.in

    Credit

    Jack and Laura Dangermond; Esri

  3. 1999

    Army software goes open

    The US Army Corps of Engineers' Construction Engineering Research Laboratory (CERL) began GRASS (Geographic Resources Analysis Support System) in 1982 to manage military lands, because no package then on sale met its needs. CERL made its last release, version 4.1, in 1992 and stopped issuing patches after 1995. In 1997, after two years of uncertainty, Baylor University in Texas and then the University of Hannover in Germany, led by Markus Neteler, took over development, and in October 1999 GRASS 5.0 came out under the GNU General Public License, which lets anyone use, change and share the code.

    LessonSoftware paid for by one agency outlived that agency's exit because others could take over its code.

    grass.osgeo.org

    Credit

    USA-CERL; Markus Neteler and the GRASS Development Team

  4. 2000

    Maps requested over the web

    The Open Geospatial Consortium (OGC), a standards body for map and location software, was founded on 25 September 1994 with eight charter members to make different vendors' systems work together. Its first standard, Simple Features, was approved in 1997, and in 2000 it approved its first web service standard, the Web Map Service (WMS): a plain web request that names map layers and an area and gets back a map image that a browser can show. OGC says it now has more than 400 members and more than 150 approved standards, all free to use.

    ogc.orgogc.org

    Credit

    the Open Geospatial Consortium and its members

  5. 2009

    Free desktop GIS goes stable

    QGIS began in February 2002 as a viewer, for Linux only, of data held in PostGIS, a spatial extension of the free PostgreSQL database that Refractions Research of Victoria, Canada, created; QGIS's first release came that June. Version 1.0, released in January 2009, ran on Unix, Linux, Mac and Windows, edited data, composed printed maps, worked with GRASS, and offered a stable programming interface for add-ons in Python or C++.

    lists.osgeo.orgwiki.osgeo.orgpostgis.net

    Credit

    Gary Sherman, who started the project, and the QGIS team

  6. 2020

    Closed license, open fork

    Mapbox GL JS is a code library for showing maps in web pages, drawn on the user's own graphics chip from vector tiles (map data cut into squares and sent as shapes, not pictures). On 8 December 2020 Mapbox released version 2.0.0 and said the library was "no longer under the 3-Clause BSD license"; upgrading meant agreeing to Mapbox's terms of service. That month users started MapLibre GL JS from the last open 1.x version, and a community organization keeps it under the BSD license, with sponsors that include Meta, which helped set up the fork.

    LessonWhen new versions of an open library closed, its users could keep the last open version going themselves.

    raw.githubusercontent.comgithub.comraw.githubusercontent.commaplibre.org

    Credit

    the MapLibre contributors; Mapbox, which wrote the original

  7. 2023

    Map files read in pieces

    A Cloud Optimized GeoTIFF (COG) is an image file laid out in tiles, with smaller overview copies inside, so that a program can fetch only the parts it needs from a web server with an HTTP range request, a request for a stretch of a file rather than the whole of it. A community of users defined it first; OGC approved it as a standard on 8 May 2023 and published it on 14 July 2023. GeoParquet, which stores map features in Parquet, a file format that stores large tables column by column, reached its first stable release, 1.0.0, on 18 September 2023.

    docs.ogc.orgogc.orggithub.com

    Credit

    the COG and GeoParquet communities; OGC

Watch for

  1. Agents that run the analysis

    Esri's documentation lists AI agents in ArcGIS Online that "perform multistep tasks, such as finding data, running analysis, and preparing results": a user describes a goal, and the agent carries out the steps and reports back. In November 2026 every listed agent was marked beta, as were most of Esri's AI assistants, including the ArcGIS Pro assistant, and the agents did not yet use credits, Esri's billing units.

    Achieved whenthe agents leave beta and an outside group publishes their success rate on a public set of GIS tasks, such as GeoAnalystBench.

    doc.esri.comarxiv.org

    Claimed by

    Esri, November 2026

Open

  1. 01

    Agents rerun GIS papers

    An AI coding agent that takes a published geographic information science paper with its shared data and code, rebuilds the computing setup, reruns the analysis and reports which figures and tables it reproduced, scored against the reproducibility reviews that people wrote for the same papers at the AGILE conference.

    The brief
    Credit · unclaimed
  2. 02

    What each format loses

    An open test suite that writes one reference dataset, with long field names, dates with times of day, non-Latin text, curved and multi-part shapes, empty values and a named coordinate system, into each common map file format (Shapefile, GeoPackage, GeoJSON, KML, FlatGeobuf, GeoParquet and Esri's File Geodatabase) through several programs, reads it back, and records every change.

    The brief
    Credit · unclaimed

Open 01 · the brief

Agents rerun GIS papers

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An AI coding agent that takes a published geographic information science paper with its shared data and code, rebuilds the computing setup, reruns the analysis and reports which figures and tables it reproduced, scored against the reproducibility reviews that people wrote for the same papers at the AGILE conference.

Why it is still open

The AGILE conference has had a reproducibility committee try to rerun accepted papers since 2020. That year it reviewed 23 accepted full papers, partly reproduced six, published its reports openly on OSF, and found insufficient documentation the main problem. Tests of AI agents in the field use single tasks: ScienceAgentBench takes 102 tasks from 44 papers in four disciplines, and GeoAnalystBench has 50 tasks scored on the workflow and code an AI model writes. ReplicationBench asks agents to replicate whole papers, but in astrophysics, and its authors found the best models scored under 20%. Koukouraki, Degbelo and Kray (2025) tested one step of a reproducibility review, comparing two map images with vision language models, and found the models can read maps but cannot yet be trusted to describe them. This research found no test of agents rerunning whole geospatial papers.

septentrio.uit.noagile-gi.euarxiv.orgarxiv.orgarxiv.orgdrops.dagstuhl.de

Built with what exists

The AGILE reproducibility reports on OSF as the answer key, and each paper's own data and code; one Docker container per paper; a coding agent driven by a model such as Claude Fable 5.2 (see the Note "Three working devices"); GDAL and the R and Python spatial packages the papers use.

  1. First step

    Download the six 2020 reports and the code of the papers they cover, and run the agent on one paper in a clean container.

  2. Done looks like

    At least 30 AGILE papers with published reproducibility reviews run by the agent, with its verdict (reproduced, partly reproduced, not reproduced) matching the human reviewer's for at least 80% of them, and every mismatch explained in a public log.

Open 02 · the brief

What each format loses

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open test suite that writes one reference dataset, with long field names, dates with times of day, non-Latin text, curved and multi-part shapes, empty values and a named coordinate system, into each common map file format (Shapefile, GeoPackage, GeoJSON, KML, FlatGeobuf, GeoParquet and Esri's File Geodatabase) through several programs, reads it back, and records every change.

Why it is still open

The limits are written down one format at a time. GDAL's documentation for its Shapefile driver says field names can be at most 10 characters, are cut and renumbered when longer, and that dates cannot hold a time of day; converter websites describe their own limits. A search found no controlled, published table of what a round trip through each format and program loses.

gdal.org

Built with what exists

GDAL and its ogr2ogr converter; QGIS; GeoPandas; DuckDB's spatial extension; ArcGIS Pro where a license is available; a Python script that compares each read-back file with the original, field by field and shape by shape.

  1. First step

    Build the reference GeoPackage with 20 deliberately awkward features and send it through GDAL's Shapefile and GeoJSON drivers and back.

  2. Done looks like

    A public table of at least seven formats by four programs, in which every recorded loss can be repeated with one command, rerun on each new GDAL release.