About the data

There is no single number for “how much is built”

This is the finding that shapes everything else, and it surprises most people. Six independent, respected products measure built-up land in the Canary Islands. They disagree by a factor of 4.7:

None of them is wrong. Each is a stricter idea of what counts as built, and they sit in a consistent order. A road is artificial, is sealed, and is not a building. A quarry is artificial and is neither sealed nor built. The 4.7× spread is the definitions, not error.

So a figure from SensiSat is only meaningful with its layer attached. This site publishes 104.33 km² of building footprint and 317.91 km² of sealed surface — both true of the same islands in the same year.

Those two differ slightly from the matching rows above, and the reason is us rather than them. The table gives each product’s own published total; these are what our layers actually hold. We rasterise the cadastre onto a 10 m grid instead of measuring its polygons exactly, and we take greenhouses out of the sealed-surface layer and publish them separately (section 7).

A seventh product, Google’s Dynamic World, reported 308 km² for Gran Canaria alone — more than double every artificial thing CORINE can find on that island. It was excluded, along with five other global datasets.

Extent and surface are different quantities

Two ways to measure the same growth, and they answer different questions:

• Extent — how much land has settlement on it. A pixel counts if it contains anything built, so it saturates: once a pixel has one house, another house does not change it.
• Surface — how many square metres are actually covered. It keeps rising as blocks fill in.

The gap between them is informative, and it reverses direction depending on how a place grows. Measured 1990 → 2015:

Each scattered new house on Fuerteventura lights up a whole 30 m pixel while adding little cover; each new block on Gran Canaria adds cover to a pixel already counted. The ratio of the two is a sprawl-versus-infill indicator.

Never compare an extent figure with a surface figure. They are different units of a different thing.

The map changes method in 2016, and says so

Before 2016 and after it, the settlement layers come from different instruments at different resolutions with different definitions. Crossing that join, on the five main islands the measured extent drops by 26–52 % — and nothing was demolished. The map simply started being made a different way. (La Gomera goes the other way: its 2016 footprint is only 3 km² and there the finer instrument finds 83 % more.)

We could have blended the two into one smooth timeline. We deliberately did not: that would bake the artefact into the data, and no reader could then tell it from a real event. The eras are published separately, the viewer announces the transition, and the measured size of the discontinuity is published alongside.

Do not add the two settlement layers together. They overlap on purpose: the 2016-onward layer’s first frame is not what was built during 2016, it is everything standing by July 2016 — the same ground the earlier layer describes, dated differently. Summed, the baseline is counted twice.

“We don’t know” is shown, not guessed

Some pixels are built, and no source can say when. In the satellite settlement layer, 37 % of the footprint that existed in 2016 carries no year at all.

There were three options: guess a year from a neighbouring pixel, assign everything to the earliest frame, or admit it. The first two would have manufactured roughly 115 km² of history that was never observed. So the data has a third state — built, year unknown — drawn in its own colour on the map, never switched on by the time slider, and hidden entirely from the change view, because a pixel with no date cannot honestly be placed in a period.

Before 1980, building dates are good to about a decade

The building layer comes from the Spanish cadastre — the official register, 474,292 buildings in the Canaries, and 98.8 % of their footprint carries a construction year. That is an unusually complete record, and it is why the pre-2016 timeline is built from it rather than from satellites.

But the register rounds dates it does not know onto round years, and it did so until 1980. Ten “bucket” years — 1900, 1910, 1920, 1930, 1940, 1950, 1960, 1970, 1975, 1980 — hold 33.8 % of every dated building in the archipelago. The year 1900 alone holds 25,971 buildings, which is 1,146 times its neighbouring years and the single largest year in the register: it is the catch-all for “old, date unknown”.

Two things stop this being worse than it sounds:

• It stops. Every bucket is 1980 or earlier. The 46 years from 1981 to 2026 contain not one, and hold 44.8 % of all dated buildings. Modern dates are good to the year.
• There is real data in between. 37 % of pre-1981 buildings carry a genuine year — the register did know the exact date for a large minority. The era is a mixture, not a set of decade stamps.

Read a pre-1980 frame as “by about then”. The viewer says so when the slider is in that range, and names the year when you land on a bucket.

Two further limits of any building register: it records what still stands, so buildings demolished before today were never registered and cannot appear; and the date is often the year of the last major reform rather than original construction. The cadastre is therefore exact for the present and a floor for the past.

Most layers cannot show anything disappear

Almost every source here is growth-only by construction: a pixel keeps the year it was first seen built, forever. That is a property of how the products are made, not a choice we made.

It costs little, because real loss is rare: measured from purpose-built change layers, built-up land disappears at 0.017–0.052 % per year. Over forty years a growth-only view misplaces roughly 2 % of built pixels.

One layer is different. The Copernicus change layers can record loss, and they did: when the Tajogaite eruption buried Todoque on La Palma in 2021, built-up land inside the lava flow went from 0.909 km² in 2018 to 0.011 km² in 2024. No other source in this project can show that, and all of them still display the buried village.

Greenhouses are counted separately, on purpose

Radar sees a plastic greenhouse roof much as it sees a building roof. Of Gran Canaria’s 27.06 km² of greenhouse parcels, the radar-based settlement product flags 72 % as built-up, while the mainly-optical one flags 9.5 %. Both are behaving correctly; they are answering different questions about the same plastic.

Left in, that would put roughly 17 km² of tomatoes into one island’s urban total. Deleted, it would lose real ground cover — under a greenhouse the soil is as covered as under a car park, which matters if your interest is habitat. So covered agriculture is removed from the urban layers and published as its own layer. Anyone who disagrees with that call can add it back exactly.

Quality varies by island — and that is a finding

This is not a disclaimer. The small western islands are measurably weaker in both of our pre-2016 sources at once:

And the satellite alternative fails on the same islands: it can date only 24.5 % of El Hierro’s settlement footprint, against 66–68 % on Gran Canaria and Fuerteventura. It maps settlement clusters, and the western islands are dispersed hamlets rather than clusters.

Work on Gran Canaria or Tenerife and both sources are solid. Work on La Gomera before 1980 and you are on the weakest ground in the archive.

How accurate is it?

Someone checked. A computer picked 911 spots at random across the islands, and a person looked at aerial photographs of each one — from 2015 and from today — and answered one question: is there a building here? They could not see our map while doing it. Then we compared. 838 of the spots gave a clear answer; the rest were genuinely too hard to call. The arithmetic follows standard practice for this kind of assessment.²

The map makes four different claims, and it is not equally good at all four:

Each check covers a 30-metre square, not a single building. So we can tell you the map puts buildings in the right place to within 30 metres. We cannot tell you it is right to within 10 m. The reason is physical: a typical building here is about 121 m², and an aerial photo shows you a roof, which sits several metres away from the walls beneath it whenever the building is tall. Below 30 metres, the photograph and the map are not describing the same thing. Our first attempt tried anyway and was thrown away.

The bottom two rows are ranges rather than single figures because the whole set was judged twice, the second time from scratch with the first answers hidden. The top two rows came out identical both times and are settled. The bottom two moved by about as much as the statistical margin, which tells you something real: those two things are genuinely hard to see on an aerial photograph, and they are the same two the reader most often could not call at all.

Before you use the time slider, read the third row. Of the squares we colour as new construction between 2015 and 2024, a quarter to a third (24–33 %) already had a building in 2015 — and that is a definition, not a fault. The Spanish cadastre resets a building’s year when it is comprehensively rebuilt: “en el supuesto de rehabilitación integral de una construcción, la fecha de finalización de dicha rehabilitación tiene la consideración de fecha de construcción”.¹ The map the cadastre itself publishes from this field is titled “Fecha de construcción o reforma integral”.

So read every year on this map as “built or comprehensively rebuilt”. A 1970 house gutted in 2019 shows as 2019, and the register is right about it — our label was what was wrong. Recent growth therefore partly shows renovation rather than new ground being taken, which matters if you are counting habitat loss.

The register misses isolated rural buildings. Of the squares we mark as empty, about 2 % hold a building — and because empty land is 95 % of the archipelago, that small percentage is the largest correction on this page. Ten of the eleven cases found have no building in the Catastro within 60 m, and seven are more than 200 m from anything mapped: they are structures in open country that the register does not hold. Some will be sheds, pens or ruins that a tax register reasonably ignores. Treat the building layer as a floor in the countryside, and as reliable in towns.

The undated class is about a quarter (28 %) empty. Where we say “holds a building, year unknown”, that share showed no building at all. Copernicus Imperviousness, an independent product the interpreter never saw, agrees on both sides of that split: a median of 91 % sealed ground where the photographs found a building, and 3 % where they did not. The leading explanation — that these are greenhouses, which radar reads as structures — was tested and is wrong: none of 107 checked points fell inside a mapped greenhouse parcel.

Corrected for both kinds of error, about 463 ± 83 km² of the archipelago sits in a 30 m square holding a building, against the 324 km² the map marks. That is an extent figure and it saturates — it is not square metres of roof, and it must not be compared with the 104.33 km² of building footprint quoted in section 1. It sits where it should between the settlement extent of satellite products (377 km²) and CORINE's artificial surface (470 km²).

One more limit, and it is the largest. Everything above is about whether buildings are there. It says much less about whether their years are right. Spain's aerial photography of the Canaries only goes back to 2005 — the islands are flown about every three years, and the older national flights never covered them — so 85 % of the buildings on this map carry a date no photograph can check. The one date boundary we could test, 2015, came back a quarter to a third (24–33 %) wrong. And there is separate evidence that the older years are worse: before 1980 the register piles buildings onto round years, with 25,971 of them dated to 1900 alone.

So: trust this map on where buildings are. Treat any year it gives before 2005 as an estimate that has not been checked and probably cannot be.

The assessment covers the cadastral building layer, archipelago-wide, at 30 m. It does not cover the satellite settlement layers, the density layers or the loss layer, and no single island has enough points for a number of its own.

Sources, licences and reuse

Every figure on this page is computed from the datasets listed above, by the methods described, and nothing here is an estimate unless it says so. The code, the evidence tables behind every number and the working notes are at github.com/LucaLovagnini/sensi-sat; each figure on this page is either generated from the published statistics or names the analysis that produced it, and a test fails the build otherwise.

How to cite

The layers, the statistics and this page are published under CC BY 4.0: reuse them for any purpose, credit SensiSat and the sources in the table above, and say if you changed them. Every source is attribution-only, which is why the whole can be too. Two conditions come from the sources: the building layer is a transformation of cadastral data and is not cadastral cartography, and the Copernicus-derived layers are modified data that the EU has not endorsed. Cite the data build you used — the figures change when the sources are re-fetched or the pipeline is re-run:

Lovagnini, L. (2026). SensiSat — urban expansion in the Canary Islands, 1900–2026 ( sensisat 0.1.0, data of 2026-09-21 ). https://sensisat.org — accessed [date]. Licence: CC BY 4.0.

For a specific number, cite the layer it comes from as well (the map names it), because the same word means different quantities across layers — see the ladder in the first section.

1. Dirección General del Catastro, Metodología — Mapas temáticos, información estadística, §5 “Fecha de construcción o reforma integral”: catastro.hacienda.gob.es. The field’s structure — one date pair per building, spanning its oldest and newest construction units — is defined in Conjunto de datos INSPIRE de la Dirección General del Catastro: English edition. SensiSat reads the oldest of the pair, which is the least renovation-affected date the register holds.
2. Accuracy assessed by the method of Olofsson et al. (2014), Good practices for estimating area and assessing accuracy of land change, doi:10.1016/j.rse.2014.02.015.