Tag: AI

Review: GeoAI

Published on by Jonathan Crowe
Book cover: GeoAI

Save some room on the AI bandwagon for ArcGIS. This seems to be the central message of GeoAI: Artificial Intelligence in GIS, a slim (only 120-page) volume of articles and posts that previously appeared, for the most part, in Esri blogs and publications. They highlight examples and “real-life stories” of how Esri’s machine- and deep-learning tools have been successfully applied in the public, private and non-profit sectors. At a moment when “AI” is invariably a synecdoche for the awfulness that is generative AI, which I will not litigate here, it can be a challenge to remember that machine and deep learning tools, which have been included in ArcGIS since 2008, have all kinds of applications and benefits. (See Esri’s pretrained deep learning models for examples like feature detection, land-cover classification, and object tracking; see also their GeoAI landing page.) Calling these tools “GeoAI” strikes me as a way to package them to appeal to decision makers who are speedrunning their AI rollout, for better or worse. It’s those decision makers that this book is targeted to. Esri has something to sell them: this is the pitch.

I received an electronic review copy from the publisher.

GeoAI: Artificial Intelligence in GIS
ed. by Ismael Chivite, Nicholas Giner and Matt Artz
Esri, 2 Sep 2025, $40
Amazon (Canada, UK), Bookshop

AllTrails and AI-Generated Hiking Trails

Published on by Jonathan Crowe

Last month the hiking app AllTrails announced AI-generated (“leveraged”) custom routes as part of a new premium membership plan, and some people are worried about it. According to the National Observer, AllTrails and other hiking apps have gotten hikers into trouble because they rely on crowdsourced trail information, which isn’t necessarily official or safe. Given generative AI’s track record for producing spectacularly erroneous results, there would appear to be some cause for concern. Except that “AI” has become a marketing buzzword that covers a lot of computer stuff, from less problematic machine learning (which is what I’d expect in this case) along with more problematic generative AI/large-language models, and AllTrails isn’t indicating which flavour they’re referring to (because: buzzword). And as the National Observer points out, “These problems already existed before the AI was added.” To be sure, generative AI is a blight on human civilization, but let’s be clear about our targets in this case.

AI Chatbots and Geolocation

Published on by Jonathan Crowe

AI chatbots don’t have the best track record when it comes to accuracy. They appear to struggle with geolocation too, as Bellingcat discovered two years ago in a test of OpenAI and Google chatbots. Bellingcat has now tested them again, this time putting 20 large-language models to work on 25 travel photos to see if things have improved, with Google Lens reverse image search as a control. The result? A few ChatGPT models outperformed Google Lens, but not by much; the rest were worse. Details at the link.

(Update: Bellingcat’s coverage goes in quite a different direction than reports last April highlighting ChatGPT’s “scary” ability to pinpoint locations from photographs, largely because it compares it with existing non-AI reverse image search. Privacy risks may not depend on the kind of technology analyzing the photo, in other words.)

Some Google Maps Updates

Published on by Jonathan Crowe

Google Maps imagery updates include improved satellite imagery thanks to an AI model that removes clouds, shadows and haze, plus “one of the biggest updates to Street View yet, with new imagery in almost 80 countries—some of which will have Street View imagery for the very first time.” The web version of Google Earth will be updated with access to more historical imagery and better project and file organization, plus a new abstract basemap layer. [PetaPixel]

Meanwhile, The Verge reports that Google Maps is cracking down on business pages that violate its policy against fake ratings and reviews.

Deep Learning Applied to Satellite Imagery Reveals Untracked Ships

Published on by Jonathan Crowe

Maps showing registered and unregistered fishing vessels near Spain, Morocco, Sicily and Tunisia.
Excerpt from Fig. 2 of Paolo et al., “Satellite mapping reveals extensive industrial activity at sea,” Nature 625 (2024).

Speaking of AI-assisted global monitoring: researchers affiliated with Global Fishing Watch have revealed that the global fishing, transport and energy fleets are a lot bigger than expected. They were able to compare the locations of ships carrying AIS transponders with satellite imagery, to which deep learning was applied to classify ships. They conclude that something like three-quarters of industrial fishing vessels, and thirty percent of transport and energy vessels, go untracked. This isn’t necessarily so much about clandestine activity—in many regions ships, especially fishing boats, simply aren’t required to be tracked—but it can, among other things, reveal illegal fishing in protected areas. Results of the study were published in Nature last month. Global Fishing Watch also has an interactive map. [The Verge]

Google, EDF Partner to Build Map of Global Methane Emissions

Published on by Jonathan Crowe

Methane is a greenhouse gas, more powerful than CO2 but shorter-lived. Google is partnering with the Environmental Defense Fund to map global methane emissions, much of which result from leaks from fossil fuel infrastructure and are undercounted. The EDF’s MethaneSAT satellite (itself a partnership between the EDF and New Zealand’s space agency) launches next month: it’ll measure methane emissions at high resolution. Google’s bringing to the party algorithms and AI, the latter to build a global map of oil and gas infrastructure.

Once we have this complete infrastructure map, we can overlay the MethaneSAT data that shows where methane is coming from. When the two maps are lined up, we can see how emissions correspond to specific infrastructure and obtain a far better understanding of the types of sources that generally contribute most to methane leaks. This information is incredibly valuable to anticipate and mitigate emissions in oil and gas infrastructure that is generally most susceptible to leaks.

More at The Verge.

Previously: Mapping Methane Emissions.

Google Maps Is Adding Generative AI

Published on by Jonathan Crowe

Uh-oh. Generative AI is coming to Google Maps. Google is using large-language models to give suggestions on where to go based on its vast horde of reviews, ratings and other contributor data. “Starting in the U.S., this early access experiment launches this week to select Local Guides, who are some of the most active and passionate members of the Maps community. Their insights and valuable feedback will help us shape this feature so we can bring it to everyone over time.” Other LLMs have a tendency to push out magnificently wrong answers; it’ll be interesting to see what results Google will get with this specific set of data. (The chances of spectacularity are not zero.)

Apple and Google Updates: AI Improvements, Airport Health Measures

Published on by Jonathan Crowe

Last week Google announced “over 100 AI-powered improvements to Google Maps” would be coming this year; these include bringing Live View indoors, a new air quality map layer, eco-friendly routing, and support for curbside pickup in business listings.

Meanwhile, Apple Maps is now displaying airport COVID-19-related health measures based on data from Airports Council International: press release. [AppleInsider, MacRumors]

Complaints about Facebook’s Automated Edits in Thailand

Published on by Jonathan Crowe

Facebook’s AI tool has added some 480,000 kilometres of previously unmapped roads in Thailand to OpenStreetMap, BBC News reports, but some local mappers have been complaining about the quality of those edits, and the overwriting of existing edits by Facebook’s editors: see OSM Forum threads here and here. In particular, see OSM contributor Russ McD’s rant on the Thai Visa Forum:

What Facebook fail to state is the inaccurate manner in which their AI mapping worked. The OSM community in Thailand had for years, been working slowly on mapping the Country, with the aim of producing a free to use and accurate map for any user. Information was added backed by a strong local knowledge, which resulted in a usable GPS navigation system based on OSM data. Main road were main roads, and jungle tracks were tracks.

Then along came Facebook with its unlimited resources and steamrollered a project in Thailand with scant regard for contributors … sure they paid lip service to us, with offers of collaboration, and contact emails … but in reality, all our comments went unanswered, or simply ignored.

Sure, their imagery identified roads we had not plotted, but along with that came the irrigation ditches, the tracks though rice paddies, driveways to private houses, and in once case, an airport runway! All went on the map as “residential roads”, leaving any GPS system free to route the user on a physical challenge to make it to their destination.

Local users commented, but the geeky humans who were checking the AI, living thousands of miles away, having never visited Thailand, just ignored our comments. They would soon move onto bigger and better things, while sticking this “success” down on their resume.

Sounds like another case of local mapping vs. armchair mapping and automated edits, where local mappers are swamped and discouraged by edits from elsewhere. [Florian Ledermann]

Previously: OpenStreetMap at the Crossroads.