Chaos Computer Club - archive feed

Machine Learning is incredibly popular at this time among researchers working with OSM data and on OSM-related problems. But what impact has this work on ML had on the OSM database or OSM community? We investigate the impact on OSM, if any, the ML work within the academic research community has had over the last few years.

# What has machine learning ever done for us?

Peter Mooney and Edgar Galvan,

Department of Computer Science,

Maynooth University, Maynooth.

Co. Kildare. Ireland.

[email protected]; [email protected]

### Introduction and background

Recently, machine learning (ML) and artificial intelligence (AI) based approaches are being applied frequently to many different types of problems in OpenStreetMap (OSM). Indeed, ML and AI have being used extensively by the research community for a plethora of applications and problems both related and unrelated to OSM. Wagstaff (2012)[1] suggests ML offers "a cornucopia of useful ways to approach problems which defy manual solutions". In specific relation to the geospatial domain, ML approaches have been reported at least as early as a decade ago with work by authors such as Werder et al. (2010)[2] on interpretation of buildings in settlements and detecting road intersections from GPS traces by Fathi and Krumm (2010)[3]. Around this time, interest in the combination of ML and OSM began to emerge. Funke et al. (2015)[4] argued that many aspects of OSM data might be suitable for "extrapolation or classification using ML". Many examples have emerged with ML approaches being used to consider problems such as: predicting or recommending tagging for objects, object classification based on contextual or proximity information, tag usage checking, automated mapping approaches, to mention some problems. Jennings et al. (2019)[5] showed that Facebook’s recent mapping campaign in OSM used ML to detect road networks from satellite imagery which are then validated by OSM editors and the local OSM communities. Examples also exist where OSM is used in ML approaches for other geospatial classification problems (Wu et al. (2020)[6], Jacobs and Mitchell (2020)[7]) while authors such as Feldmeyer et al. (2020)[8] used machine and deep learning algorithms with OSM for developing socio-economic indicators. Audebert et al. (2017) provided additional examples and argued that OSM's richness means it can be used in difficult problems such as semantic labeling of aerial and satellite images.

In addition to the observations by Vargas-Munoz et al. (2021) in their recent review of ML approaches in OSM, we can usually observe ML and OSM interaction in one of three ways: (1) ML approaches are used to improve or correct OSM data, (2) instances where OSM is used as a means of training ML models for some specific task such as building segmentation, road speed estimation (Keller et al, 2020 [10]) or land use classification (Schultz et al. 2017 [9]) or (3) where the contribution patterns of OSM contributors are analysed using ML techniques as in work such as that by Jacobs and Mitchell (2020)[7]. In this submission we ask the following question. With all of the many applications and integration of ML and AI with OSM, over the past number of years, how many of these applications and approaches have been adopted or used by the OSM community? Furthermore, what are the benefits or impact of these efforts from the research community with ML and AI approaches to the OSM project and OSM community? We believe that there is significant scope for ML researchers to make impactful and helful contributions directly within OSM on problems such as tag updating and correction, added intelligence within OSM editing software, intrinsic quality analysis, etc.

### Methodology and Findings achieved

A systematic review of approximately 60 peer-reviewed academic journal and conference papers will be reported. These papers are selected on the following basis that the paper(s): (1) clearly outlines an ML or AI approach using OSM data, (2) tackle a problem known in the OSM community such as tag prediction, contribution patterns, or geometry correction. Paper metadata such as title, keywords, and abstract contents are used to select the papers. Manual checking of the papers is also undertaken to ensure that the content of each paper relates to our selection criteria. A classification of these papers will be developed based on the following set of questions:

* What are the most common ML approaches used by researchers for the three instances outlined above? For example, Learning Problems (supervised, self-supervised, reinforcement), Statistical Inference (Inductive, Deductive), etc.

* What are the most common types of problems in OSM tackled by ML approaches? For example, automated tagging, contribution pattern analysis, intrinsic quality analysis, object classification, etc.

* Are the approaches reproducible and replicable for other regions or areas within OSM? For example, is a particular ML approach limited to a specific geographical area or thematic area (such as roads, buildings, waterways, etc.) in OSM.

On this classification we then report a narrative on our findings on the benefits and impacts of these efforts to the OSM project and the OSM community. We are working on this analysis at the time of writing.

### Final Discussion of scientific contributions

As suggested by Jacobs and Mitchell (2020)[7], ML can "contribute to the diversification and quality of available assessment methods for OSM" while Feldmeyer et al. (2020)[8] argues that the application of ML to OSM can reveal the "untapped potential for knowledge generation" in OSM. In our work, we argue that we must not get carried away with the combination of ML and OSM purely for the sake of it. OSM, as a massive open geospatial database, is a very attractive source of (geo-)data for researchers and practitioners looking to train, benchmark and test ML approaches. Consequently, we can confidently state that, after well over a decade of reported results in this domain, researchers have produced many excellent research and knowledge outputs using the ML and OSM combination. Now we enter a phase of technological and scientific development with ML and OSM where we must ask how can all of this ML knowledge contribute effectively to the OSM database and OSM community.

Grinberger et al. (2019)[11] argue that while efforts to establish and strengthen interaction between the research community interested working with or in OSM and the OSM community itself have generally been positive. However, opportunities exist to enhance interactions between these two communities and perhaps ML could be the catalyst for a new interaction. Based on this the scientific contribution of this work is multi-faceted. Firstly, this paper will stimulate debate about the contribution of these ML approaches to the improvement of OSM data and enhancement of the OSM community. Secondly, this work will highlight situations where these ML approaches have delivered genuinely new and novel outputs of interest to OSM in general. Finally, this work will issue the challenge to the academic community to apply ML to several interesting and open problems which are of mutual interest to both the academic and OSM community.

about this event: https://2021.stateofthemap.org/sessions/RHR7Q8/

NLMaps Web is a web interface for querying OSM with natural language questions such as “Show me where I can find drinking water within 500m of the Louvre in Paris”. They are first parsed into a custom query language, which is then used to retrieve the answer by queries to Nominatim and Overpass.

Nominatim and Overpass are powerful ways of querying OSM, but the Overpass Query

Language is somewhat impractical for quick queries for unfamiliar users. In

order to query OSM using natural language (NL) queries such as “Show me where I

can find drinking water within 500m of the Louvre in Paris”, Lawrence and

Riezler [1] created the first NLMaps dataset mapping NL queries to a custom

machine-readable language (MRL), which can then be used to retrieve the answer

from OSM via a combination of queries to Nominatim and Overpass. They extended

their dataset in a subsequent work by auto-generating synthetic queries from a

table mapping NL terms to OSM tags – calling the combined dataset NLMaps v2. [2]

The proposed purpose of these datasets is training a parser that can parse NL

queries into their MRL representation, as done in [2-5].

The main aim of my Master’s thesis was building a web-based NLMaps interface

that can be used to issue queries and to view the result. In addition, the web

interface should enable the user to give feedback on the returned, either by

simply marking the parser-produced MRL query as correct or incorrect, or by

explicitly correcting it with the help of a web form. This feedback should be

directly used to improve the parser by training it in an asynchronous online

learning procedure.

After observing that parsers trained on NLMaps v2 perform poorly on new queries,

an investigation into the causes for this revealed several shortcomings in

NLMaps v2, mainly: (1) Train and test split are extremely similar limiting the

informativeness of evaluating on the test split. (2) Various inconsistencies

exist mapping from NL terms to OSM tags (e.g. “forest” sometimes mapping to

natural=wood, sometimes to landuse=forest). (3) The NL queries’ linguistic

diversity is limited since most of them were generated with a very simple

templating procedure, which leads to parsers trained on the data not being very

robust to new wordings of a query. (4) In a similar vein, there is only a small

amount of different area names in NLMaps v2 with the names “Paris”, “Heidelberg”

and “Edinburgh” being so dominant that parsers are biased towards producing

them. (5) Some generated NL queries are worded very unnaturally making them

counter-productive learning examples. (6) Usage of OSM tags is sometimes

incorrect, which affects the usefulness of produced parses.

The detailed analysis is used to eliminate some of the shortcomings – such as

incorrect tag usage – from NLMaps v2. Additionally, a new approach of

auto-generating NL-MRL pairs with probabilistic templates is used to create a

dataset of synthetic queries that features a significantly higher linguistic

diversity and a large set of different area names. The combination of the

improved NLMaps v2 and the new synthetic queries is called NLMaps v3.

A character-based GRU encoder-decoder model with attention [6] is used for

parsing NL queries into MRL queries using the configuration that performed best

in previous work [5]. This model is trained on NLMaps v3 and used as the parser

in the newly developed web interface. Mainly through advertising on the OSM talk

list and the OSM subreddit, 12 annotators are hired from all over the world to

use the web interface to issue new NL queries and to correct the parser-produced

MRL query if it is incorrect. They are assisted by completing a tutorial before

the annotation job and by help compiled from taginfo [7], TagFinder [8] and

custom suggestions for difficult tag combinations. The collected dataset

contains 3773 NL-MRL pairs and is called NLMaps v4.

With the help of NLMaps v4, an informative evaluation can be performed revealing

that a parser trained on NLMaps v2 parses achieve an exact match accuracy of

5.2 % on the MRL queries of the test split of NLMaps v4 while a parser trained

on NLMaps v3 performs significantly better with 28.9 %. Pre-training on

NLMaps v3 and fine-tuning on NLMaps v4 achieves an accuracy of 58.8 %.

Since the thesis’s goal is an online learning system – i.e. a system that

updates the parser directly after receiving feedback in the form of an NL-MRL

pair –, various online learning simulations are conducted in order to find the

best setup. In all cases, the parser is pre-trained on NLMaps v3 and then

receives the NL-MRL pairs in NLMaps v4 one by one, updating the model after each

step. The most simple variant of the experiment uses only the one NL-MRL pair

for the update, another variant adds NL-MRL pairs from NLMaps v3 to the

minibatch and a third variant additionally adds further “memorized” NL-MRL pairs

from previously given feedback to the minibatch. The main findings of the

simulation are that all variants improve performance on NLMaps v4 with respect

to the pre-trained parser, but with some of them the performance on NLMaps v3

degrades. The simple variant that updates only on the one NL-MRL pair is

paricularly unstable, while adding NLMaps v3 instances stabilizes the

performance on NLMaps v3 and improves the performance on NLMaps v4. Adding the

instances from memorized feedback further improves the performance to an

accuracy of 53.0 %, which is still lower than the offline batch learning

fine-tuning mentioned in the previous paragraph.

In conclusion, the thesis improves the existing NLMaps dataset and contributes

two new datasets – one of which is especially valuable since it consists of real

user queries – laying the groundwork necessary for further enhancing NLMaps

parsers. The current parser – achieving an accuracy of 58.8 % – can be used by

OSM users via the new web interface currently available at

https://nlmaps.gorgor.de/ for issuing queries and also for correcting incorrect

ones. Future work will concentrate on improving the web interface’s UX and

enhancing the parser’s performance in terms of speed and accuracy.

about this event: https://2021.stateofthemap.org/sessions/GDMBWS/

Between December 2020 to March 2021, the OSM Kenya team ran a project whose focus was on growing the OSM community locally: both in terms of membership and diversity. The project was supported through the Facebook and HOTOSM community impact microgrants. In this session, we will share about the community, project: from ideation(community health), motivation to implementation. This will include the activities involved, our experience, challenges that we encountered, and the lessons learned.

Being one of the awardees for the 2020 Community Impact Microgrants by HOTOSM and Facebook, the OpenStreetMap community in Kenya ran a virtual three-month training program for women and girls interested in OSM. While the project was implemented towards the end of 2020, the ideation process began at the start of the year when discussing community health and brainstorming on activities to conduct virtually due to COVID-19 restrictions. This included carrying out a community survey, identifying gaps and challenges, and figuring out what next. Our key focus was community growth, especially in terms of diversity and membership. We had some virtual activities that introduced new mappers to the community but still saw no diversity improvement.

We, therefore, designed a women-only training program that would run for eight weeks. We were able to train a group of 35 beginner-level and intermediate-level mappers on all aspects of open mapping. Activities included hands-on workshops, presentations, and mapathons. During this period, we had some successes, learned a lot, but we also had a good share of challenges. We hope that by sharing our experience, other communities interested in implementing similar programs can learn more about the same; we also hope to have discussions that would improve future iterations of the program.

about this event: https://2021.stateofthemap.org/sessions/L9XART/

Mappers see the OpenStreetMap Standard layer every time they view OpenStreetMap.org, but who else is using the layer? With usage logs, who is accessing what is broken down for this important OpenStreetMap service.

The OpenStreetMap Standard layer is one of the high profile OpenStreetMap services, but not many people know who uses it, and where. This talk covers the goals of the service, it’s history, how it's designed, and dives deep into how it's used. Using improved logging functionality, we’ll get answers to these questions and more.

We’ll cover the policy and technical basics of the service, showing how changes make it to your browser through the various elements of the stack and what we set out to do by running the standard layer, including policy and history.

Diving deep into the usage, we’ll look at where people are viewing, where they’re from, daily and weekly usage patterns, and then into what sites and apps people use to view the layer, including the long tail of small sites. You’ll see how different uses put different loads on the OSMF servers, and how load doesn’t come from where people think it is.

Many people just want to see their edits on the map, and we’ll cover that too, looking at how they show up and what to do if they don’t.

Along the way we’ll learn how to see your changes on the map and see heatmaps, treemaps, and find out just what’s up with the teapots.

about this event: https://2021.stateofthemap.org/sessions/GGLYCK/

Pedestrians have been neglected. We’ve seen monumental progress in digital maps, but much of this has been road centric. In this presentation we download data from OpenStreetMap that relates to pedestrians to see how much it differs from the reality on the ground. We contrast different types of cities, seek to understand why pedestrian data is lacking, and look at solutions such as Mapillary that can help make OpenStreetMap more pedestrian friendly.

The evolution of digital maps of the last 20 years has been nothing short of incredible. The experience for the end consumer has continued to improve, with better map data, more intuitive interfaces, and greater portability. A lot of the developments have focused on in-car navigation, with Google Maps, Apple Maps, HERE Maps, and TomTom dedicating significant resources to the space. Even in our beloved OpenStreetMap, vehicle based navigation claims centre stage.

In this presentation we’ll explore the state of pedestrian data in OpenStreetMap, how it differs between cities, why it’s important to think about, and how we might collectively improve the quality of pedestrian data.

To begin with, we’ll take a look at data downloaded with Overpass Turbo. The data represents nodes, ways, and areas with pedestrian relevant tags such as highway=footway and sidewalk=both. Our analysis focused on five cities with differing characteristics:

- Folsom, USA

- Heidelberg, Germany

- Melbourne, Australia

- Stone Town, Tanzania

- Yesan, South Korea

These cities differ in population, cultural characteristics, urban planning, history, and topography. We’ll explore what kind of OpenStreetMap tags have been used in each city, how close this matches the state of pedestrian infrastructure, and how the cities compare to one another.

We’ll then look at some reasons why pedestrian data has been neglected including the limitations of satellite imagery, commercial incentives, and data collection methods. In the final part of the presentation we’ll propose and hopefully discuss some of the tools that could help including Mapillary, StreetComplete, GoMap!!, and Vespucci.

about this event: https://2021.stateofthemap.org/sessions/RFGVSG/

The code samples, status, recent development and roadmap of the open-source community driven project for rendering of (not only) OpenStreetMap vector tiles in a web browser (GL JS) and with native code (Android, iOS, etc).

After Mapbox announced the closure of Mapbox GL JS, their JavaScript library for displaying maps using WebGL, the community around Hacker News gathered on Slack and GitHub and made a collective decision to maintain and further develop the last open-source version of the software and build a 100% free alternative of the project. This is how the MapLibre was born.

As a group of individuals, we coordinate the effort and synchronize contributions from multiple teams (MapTiler, Amazon, Facebook, Elastic, Stadia, Microsoft, Jawg, GraphHopper, Toursprung, etc) - working on JavaScript and Native code implementation of the renderers and related ecosystem.

Multiple releases have been published, the project has CI checks for contribution, regular steering committee meetings, updated support for TypeScript, several bindings such as ReactJS, the Metal rendering on iOS is implemented (as Apple decided to deprecate OpenGL ES), and many issues and bugs has been fixed. There is plenty of ideas what to do next - from implementation of 3D terrain rendering, to support of non-Mercator map projections, or tighter integration with Leaflet, and much more.

Let's explore the current status of the project, learn how to use MapLibre in your own software with practical code samples, and how to join and contribute to the collaborative development and participate on a shared roadmap.

about this event: https://2021.stateofthemap.org/sessions/RC7S8K/

Since moving to Ireland from Germany in December 2016, I have been mapping heritage features on the island increasingly. During lockdown, I started a YouTube channel to enable other people who are interested in Irish history to learn about OpenStreetmap, because it is still quite unknown in those circles.

After moving to Ireland and shortly after becoming a member of a local historical society, I started mapping historical features like ringforts, castles, church ruins and graveyards. As a member of that group, I noticed that the awareness of OpenStreetMap and OpenData is very low in this group and in Ireland in general. One of the excuses often given by these groups is the demographic which is less computer literate than my generation. However, after uploading a few OSM tutorials onto YouTube, I received good feedback and two female acquaintances of that age group started actively contributing to OSM. Motivated by those developments and by looking for something productive to do during the third COVID lockdown, I started a YouTube tutorial series on OSM, uMap, overpass turbo and fieldpapers to show the benefits of OSM to historically interested individuals and groups. Within only a few weeks, I had gathered a small following and continue getting positive feedback from people who discover OpenStreetMap for their personal or group projects. I believe that giving the public concrete application examples of OSM rather than trying to convince them by dropping terms like OpenData and OpenSource into the conversation, is a way to promote OpenStreetMap to an audience that is keen to learn and to apply the knowledge. It is a time consuming process, but necessary for a shift in awareness about OpenStreetmap.

about this event: https://2021.stateofthemap.org/sessions/ZCUPCF/

The osm2pgsql has been around for a long time. Since 2006 it is used to import OSM data into PostgreSQL/PostGIS databases for rendering and other uses. In the last years there have been a lot of improvements to osm2pgsql. With the new "Flex" output, osm2pgsql is now much more versatile which allows new applications.

The osm2pgsql has been around for a long time. Since 2006 it is used to import OSM data into PostgreSQL/PostGIS databases for rendering and other uses. In the last years there have been a lot of improvements to osm2pgsql. With the new "Flex" output, osm2pgsql is now much more versatile which allows new applications.

The talk will be about the new developments in osm2pgsql und show how to use the new features. Osm2pgsql is used for rendering bitmap tiles for many maps, including the main OSM map and many others. But it can also be used to import OSM data for generating vector tiles and for many other use cases. It is now possible to tell osm2pgsql exactly what output tables with what fields to create in the database. Transformations from the OSM tags to more or less any format you need in your database can be defined in Lua code. Together with the powerful SQL query language in PostgreSQL and the geometric algorithms in the PostGIS plugin, this makes all sorts of analytical processing possible. Osm2pgsql can be used to import small OSM extracts quickly for one-off processing or run a minutely updated planet-wide database on a reasonably-sized machine making OSM data processing accessible for everybody.

about this event: https://2021.stateofthemap.org/sessions/7RWSWA/

A deep dive into how StreetComplete stores edits and syncs them with OpenStreetMap, including persistence, solving conflicts and allowing users to revert their edits.

StreetComplete is known as an app to easily contribute selected data to OSM on the go.

The simple interface could make it appear that the technical implementation is equally simple. It is maybe a common misconception that things that look easy to the user are also lighter in terms of code complexity. But under the hood, it is anything but.

For example, amongst other things, the app has a unique way to avoid and to automatically resolve conflicts when uploading data to OSM. For an app that promises its users to work completely offline and thus automatically stores unsynced edits for any duration before upload, this is very useful to have.

Furthermore, it comes with the ability to undo edits in any order (not just the last) and even undo most edits after they have been synced with OSM already.

This talk shall give you an architectural overview over how StreetComplete enables its users to do these things.

It could serve other editor developers as inspiration, however, most technical concepts in this app are probably not well applicable to be used in a general OSM editor, as StreetComplete vastly limits its users what they can do and therefor we can make certain assumptions about how the app is used.

about this event: https://2021.stateofthemap.org/sessions/7M7WVX/

This track gathers various lightning talks of 5 minutes each.

## The Boundary Puzzle

*ark Arjun* | *[arkarjun](https://www.openstreetmap.org/user/arkarjun)*

The story of OSM Kerala community in developing the administrative boundaries of Kerala, a southern state of India. The Mammoth effort of volunteers made the different level of administrative boundaries of the state created a large impact in recent times including the pandemic management. The talk gives a brief on how the citizen-oriented participatory mapping revolutionised the open data culture and influenced the authorities.

## The YouTube community of OpenStreetMappers

*Gregory Marler* | *[LivingWithDragons](https://www.openstreetmap.org/user/LivingWithDragons)*

Let's take a look at YouTube and the community there from an OpenStreetMap view. I quickly cover the big and important channels that host talks, provide tutorials and learning, along with those doing desktop streams or turning on a camera and chatting as they go out mapping.

## Let's meet at an island. OpenStreetMap as a source for spatial chat tool worlds

*Helga Tauscher*

This talk shows how worlds for adhoc video-conferencing systems or spatial chat tools, for example [WorkAdventure](https://workadventu.re/) can be populated from geospatial or other real-world data on an automated basis. The island scenario is the first of three scenarios to be investigated, with the other two being scenarios on the building and city district level.

## Experiments in P2P tiles

*Iván Sánchez Ortega* | *[ivansanchez](https://www.openstreetmap.org/user/ivansanchez)*

An exploration on how a seemingly obscure web standard (WebRTC) can be exploited to provide P2P transmission of rendered map tiles and theoretically lower the load on the OSMF's tileservers.

## Open Healthcare Access Map

*Marcel Reinmuth* | *[maze2p0](https://www.openstreetmap.org/user/maze2p0)*

Get insights on travel time to healthcare facilities, population coverage and their interplay at different scales for a variety of countries. The [Open Healthcare Access Map](https://apps.heigit.org/healthcare_access/#/) is an OSM data powered application to explore health access.

## Bye Bye, Unclassified

*Martijn van Exel* | *[mvexel](https://www.openstreetmap.org/user/mvexel)*

Many words have been said and written about road classification in OSM. In this talk, Martijn van Exel will look at this topic through the lens of recent community initiatives in the United States, and add his own opinions you never asked for.

## Project OsmAPP.org

*Pavel Zbytovský* | *[zby-cz](https://www.openstreetmap.org/user/zby-cz)*

For a long time, I missed the single application for using OpenStreetMap. In this talk I would like to present the universal OpenStreetMap app, OsmAPP for short. After a few years of development, it already offers fast vector maps, clickable POIs, search and even basic editing capabilities. Let's hear the story of this project and discuss the future.

- Link: [osmapp.org](https://osmapp.org)

- Slides: [github.com/zbycz/osmapp-talk](https://github.com/zbycz/osmapp-talk)

about this event: https://2021.stateofthemap.org/sessions/ZLK9X9/