narrobotR (v2)

Cruising the English and Welsh canals on Bluesky and Mastodon.

🦋 Bluesky: https://bsky.app/profile/narrowbotr.bsky.social
🦣 Mastodon: https://mastodon.social/@narrowbotr

The narrowbotR (pronounced "narrow-bot-r") is a bot on Bluesky Mastodon that posts random locations on the Canal & River Trust (CRT) network. Where possible it sources a photo of the location from Flickr, otherwise it posts an aerial photograph sourced from Mapbox. The bot is written in R and powered by Github Actions.

Background

This is the second edition of the bot. The first version of the bot originally ran on Twitter (now X) from April 2020 but this was discontinued in April 2023 following changes to Twitter's terms of service. The bot also ran on Mastodon from November 2022 and on Bluesky from December 2024. In May 2025, the first edition of the bot suffered from repeated failures of the Github Action to post content to Bluesky. Despite numerous attempts to identify the source of the failure, this was unsuccessful and so the bot was taken offline.

This second edition is a fresh recreation of the bot, with some adjustments to the code, and rebuilding of the database. The original bot's code remains available in the original narrowbotR repo.

CRT locations

The bot selects locations from a database of places derived from the Canal & River Trust's open data feeds. There are 17,144 points included in the dataset representing 14 different types of feature within the English and Welsh canal network. The first edition of the bot used only point based data feeds, in this second edition the "canals by km" feed has also been used to include sections of canal. For these canal sections the centroid of each segment has been included in the data. This should allow for potential photographs on or by canal towpaths that are not otherwise near a feature such as a lock or a bridge.

Some 40% of the features in the database are bridges while locks, which are synonymous with canals only account for 10% of places. Similarly culverts, which are often hidden from direct view, account for 17% of the places included in the dataset but boat lifts, dry docks, lakes, slipways, tunnel portals and wharves all account for only around 2% in total. These features are more likely to give rise to options for interesting photography than features such as culverts and weirs (most weirs on the CRT network, like culverts are often out of direct view). While bridges may give rise to photographic options due to their large presence in the dataset these have been down-weighted, as have the centroids of canal sections. Within the bridges set of features, bridges identified as "pipe bridges" (i.e. those containing services such as mains water or sewerage) have been further down-weighted than bridges carrying roads, track, footpaths or railways.

In the weighted dataset, locks and bridges each now have around a 1 in 3 chance of being selected and all features have at least a 0.5% chance of being picked. As a result, in a given year the bot might reasonably be expected to select the Anderton Boat Lift around 7 times, whereas in the unweighted dataset it was likely to be selected only once every 11 years. The table below shows the raw and weighted counts and percentages of features in the database, it also includes a count of the likely sightings of different feature types in a given year (assuming 365 days and four posts per day).

Feature type	Raw #	Raw %	Weighted #	Weighted %	Likely sightings
lock	1,722	10.0%	3,444	33.0%	481
bridge	6,916	40.3%	3,330	31.9%	465
canal_km	3,173	18.5%	794	7.6%	111
aqueduct	330	1.9%	660	6.3%	92
winding_hole	533	3.1%	533	5.1%	74
tunnel_portal	103	0.6%	412	3.9%	58
wharf	75	0.4%	300	2.9%	42
dry_dock	68	0.4%	272	2.6%	38
pumping_station	88	0.5%	176	1.7%	25
weir	1,108	6.5%	166	1.6%	23
culvert	2,960	17.3%	148	1.4%	21
slipway	57	0.3%	114	1.1%	16
boat_lift	1	0.0%	50	0.5%	7
lake_pond_fishery	10	0.1%	50	0.5%	7

Workflow

The narrowbotR runs via Github Actions four times per day. The bot's workflow is as follows:

1. Select a place at random from the locations database.
2. Use the Flickr API to identify nearby photos (if any). a. If Flickr photos exist then give each photo a score and select the highest scoring photo. b. If no Flickr photos exist then generate a satellite image of the place from the Mapbox API.
3. Download the photo from Flickr/Mapbox.
4. Construct the text of the post: a. The place name (including if available the waterway name). b. A link to the location on OpenStreetMap. c. If a Flickr photo, attribution and a link to the Flickr photo page. d. Tags including the base tags of #canal, #narrowboat, either of #england or #wales depending on location, and # uk. For flickr photos any tags included by the Flickr user will be appended and for mapbox photos a set of tags indicating it is arial photography. Additional tags are limited to ensure total post length does not go over 300 characters.
5. Attempt to post on Mastodon and Bluesky.

CRT locations database

The database of CRT locations is derived from the CRT open data feeds (see feeds.csv). The database is generated offline since it would be computationally expensive to recreate each time the bot is run, and unnecessary since there is only limited change in the CRT's features over time.

The database combines locations provided from point data as well as including the centroids computed from line segments representing 1 kilometre stretches of canal and waterway centrelines. The feature sub-type is also included included in the database (e.g. to identify which bridges are road bridges vs railway bridges), at present this is only used for weighting purposes.

Welsh marker

Around 4% of the places in the database are located within Wales, however this is not marked in the CRT data feeds (the canal segments data feed includes a region marker but Wales is part of a combined region with South West England).

In constructing the database a marker for places in Wales is included so that a #wales hashtag can be included in posts rather than #england.

Flickr photo scoring

The bot tries to source photos from the Flickr API that are within a 100 metre radius of the selected location's latitude and longitude. Only photos that have some form of publicly reusable license are selected (i.e. excluding photos that are set to "All rights reserved").

The bot calculates each score using four measures:

• time_score: photos taken within "golden hour" are given a score of 2, those in standard daylight a score of 1.5, those during twilight a score of 1, and those at night a score of 0.5.
• recency_score: photos are given a score between 1 and 2 depending on how recently the photo has been taken, with more recent photos given precedence over older photos.
• distance_score: no location information is returned from the Flickr API search, however based on testing of the API it appears that photos are returned sorted by distance. A pseudo-score of between 1 and 2 is generated based on the photo's position in the search results, with the first photo getting a score of 2 and the last photo getting a score of 1, all other photos are scaled evenly based on their position in the results list.
• word_score: the photo's title, description and tags are assessed for their use of 94 words that relate to either canals/waterways, wildlife and other landscape terms. The word score also ranges from 1 to 2, but is squared to increase its effect in the final calculations.

Golden hour and solar calculations

The previous version of the narrowbotR used the {suncalc} package to determine times such as sunrise, sunset and the morning and evening golden hours. To reduce dependencies these are now calculated within the repo based on formulae published by the US National Oceanic and Atmospheric Administration (see solareqns.PDF¹).

There are several definitions of "golden hour", the time immediately after sunrise or immediately before sunset where, due to the angle of the sun, natural light is more red and creates golden hues. As a result "golden hour" is often a prized time for photography. For the narrowbotR, the opposite of the civil twilight period has been used, i.e. when the sun is within 6º of the horizon.

Wordlist

The bot checks Flickr photos for the usage of the following words in the title, description or tags.

Canal related terms

aqueduct	bank	barge	beam	boat
bridge	canal	channel	cut	cruise
cruising	dinghy	flight	embankment	gate
gongoozl	junction	keeper	lock	marina
mooring	narrow	nav	paddle	path
piling	pound	quay	quayside	river
sluice	tow	tunnel	water	weir
winding	windlass

Wildlife related terms

badger	bat	bee	bird	butterfly
coot	comorant	cow	damselfly	dormouse
dragonfly	duck	fish	frog	fox
goat	goose	grasshopper	grass snake	heron
kestrel	kingfisher	mallard	newt	otter
owl	polecat	sheep	stoat	swan
vole

Landscape and heritage related terms

cargo	coal	cottage	industr	glade
heritage	historic	house	lake	landscape
factory	field	flower	forest	meadow
mill	park	pond	pub	pump
reservoir	textile	transport	tree	warehouse
wood

License

The code in this repository is licensed under the MIT License, this covers all R scripts in the R folder, the Github Actions workflows and the supporting documentation in this README.

The original location data is source from Canal & River Trust's Open Data service. Different data feeds are subject to different license conditions, with one of three different licenses applying. The feeds.csv file provides the license conditions for each individual feed:

• The Open Government Licence (OGL): https://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/
• The Canal & River Trust's Data License (CRTDL) (PDF, web)
• The Ordnance Survey's INSPIRE License (INSPIRE) (PDF, web)

As far as is known, any data requiring the Ordnance Survey INSPIRE license is not used by the bot.

The source data remains copyright of the Canal & River Trust: © Canal & River Trust copyright and database rights reserved 2026.

Footnotes

1. The NOAA host an online calculator, included within the calcuator's links page is a PDF providing formulae for general calculation of sunset and sunrise. As the calculator is no longer actively maintained the PDF has been downloaded and stored in the ref folder of this repo. ↩