Every list of free elevation data has the same problem: it was true once. Links rot, portals move behind a login, and a 2019 tutorial still points at a directory that 404s.

The short answer, if you only want one: Copernicus GLO-30 for a global surface model, NASADEM for bare earth, and USGS 3DEP if your area of interest is in the United States. The rest of this page is what each one costs you in licence, vertical datum and file size — and which of them a browser can actually open.

So before writing this we probed 48 URLs across 16 sources on 11 September 2026: eight open with a cross-origin header, twelve open as downloads without one, fourteen answering without range support, two wanting an account, three refusing our script, five redirecting, four dead.

Each URL got one GET with a Range: bytes=0-3 header, an Origin header and a browser user agent, redirects not followed; we read the status, accept-ranges, access-control-allow-origin, the content-range total and the content-type.

Two lessons. A 206 does not mean you found a raster — six pages we probed are HTML on a server that supports byte ranges. And a probe is one moment, one IP, one user agent: a host can refuse a script and serve a real browser, so a 403 is sometimes a fact about the probe, not the source.

DEM, DSM and DTM are not three words for the same thing

A DEM is the umbrella term: a raster where each cell holds a height. What that height is of is what matters. A DSM records the top of whatever is there — canopy, rooftops, a bridge deck. A DTM, or bare earth, records the ground underneath, and over forest the two differ by tens of metres.

The most-recommended modern global dataset, Copernicus GLO-30, is a DSM, and so is ALOS AW3D30. Compute a flood surface from one and your answer is wrong wherever there are trees and buildings.

The vertical datum trap, or why two 30 m DEMs disagree by 30 metres

Two DEMs can cover the same point, at the same resolution and CRS, and report heights tens of metres apart. Usually nothing is broken — they are measured from different surfaces. An ellipsoidal height comes from the WGS 84 ellipsoid, an orthometric height from a geoid following mean sea level, and the gap runs from about −100 m to +85 m.

The second trap is that two orthometric DEMs can use different geoid models. The SRTM family uses EGM96, EPSG code 5773; Copernicus GLO-30 uses EGM2008, EPSG code 3855. The two differ over one point by the gap between those geoids plus the surface-versus-bare-earth difference, and neither is an error. Each entry states its own; to compare two, transform first, using our EPSG cheat sheet for Brazil and Latin America.

The global sources

SRTM — the baseline everybody starts from

The 2000 Shuttle Radar Topography Mission is still the reference, in two resolutions.

Resolution — 1 arc-second, about 30 m at the equator, for SRTMGL1 v003; measured pixel 0.000277777777778 degrees. The void-filled CGIAR-CSI product is 3 arc-second, 90 m, vertical error under 16 m.
CRS and vertical datum — WGS 84, EPSG:4326, read from the file header; EGM96 geoid, EPSG:5773, orthometric, per OpenTopography's dataset page.
Coverage — global land from 60 degrees north to 56 degrees south, acquired February 2000.
Licence — NASA and USGS data, free to use; cite the DOI 10.5067/MEASURES/SRTM/SRTMGL1.003. For the 90 m product we could not verify one: the CGIAR-CSI page carried no licence or terms statement, so read theirs before redistributing.
How to get it — the earthdata catalogue entry is the portal; OpenTopography's São Paulo tile is a file, 15,340,529 bytes, three overview levels. NASA's own 3 arc-second tiles — not CGIAR's void-filled build — are mirrored at srtm.kurviger.de as HGT zips.
Verdict — open; the OpenTopography copy is readable by a browser application, the 90 m mirror is not.

Copernicus DEM GLO-30 — the best modern global surface model

If you want one global 30 m dataset in 2026 and do not need bare earth, this is it.

Resolution — 1 arc-second, about 30 m. The readme notes the number in COG_10 is arc seconds, not metres — hence a 30 m product labelled 10.
CRS and vertical datum — WGS 84, EPSG:4326, measured; EGM2008 geoid, EPSG:3855.
Coverage — global, 2011 to 2015, a DSM. Beyond 50 degrees north or south, ESA's native grid and OpenTopography's resampled copy differ.
Licence — its own terms, not public domain: free for the general public under the Copernicus licence, per the AWS registry entry and the bucket readme.
How to get it — the tile index lists every key and the São Paulo tile is 42,276,212 bytes. The STAC collection resolves a bounding box to the right tile but returns an s3:// URI no browser can open, so use it to learn which tile, then rewrite the key for the HTTPS host.
Verdict — open as a download, not readable by a browser application: no access-control-allow-origin header at all, and pasted into a viewer it is refused at the CORS preflight with zero bytes on the wire.

NASADEM — SRTM, reprocessed

NASADEM reprocesses the original radar with better orbit data and fills voids from ASTER GDEM and ALOS PRISM. If you were going to use SRTM, use this instead.

Resolution — 1 arc-second, about 30 m, same measured pixel size as SRTM.
CRS and vertical datum — WGS 84, EPSG:4326, measured; EGM96 geoid, EPSG:5773, per OpenTopography's dataset page.
Coverage — SRTM's band but larger: 147,385,570 square kilometres against SRTM's 119,560,000, because void filling extended the footprint.
Licence — NASA data, use licence recorded as not provided, so cite the DOI 10.5067/MEASURES/NASADEM/NASADEM_HGT.001.
How to get it — the catalogue entry is the portal; the bare-earth São Paulo tile is a file, 15,067,744 bytes.
Verdict — open, and the OpenTopography copy is readable by a browser application.

ALOS World 3D-30m — free, but the front door asks for an account

A good independent check on the SRTM family, with the most misreported access story here.

Resolution — 1 arc-second, which JAXA describes as approximately 30 metres, latitude dependent.
CRS — geographic latitude and longitude on WGS 84 per the product description; not measurable here, the objects being zips.
Vertical datum — height above sea level, so geoid-referenced, and the page does not name the geoid model. Plenty of posts say EGM96; we could not verify it, so we do not.
Coverage — global, 23,993 tiles of one degree each. It is a DSM.
Licence — free of charge for commercial and non-commercial use under the terms on the product page.
How to get it — the download entry point returns 401 and wants an account; the release tiles do not, and one tile on the Rio de Janeiro coast answered 206 at 356,019 bytes.
Verdict — registration is true of the entry point only, not the objects. Once you know a key it is an open download, but not one a browser application can read.

GEBCO — the one that covers the ocean floor

Every other source here stops at the coastline. GEBCO does not.

Resolution — 15 arc-second, roughly 450 m at the equator; measured 0.004166666666667 degrees per pixel.
CRS and vertical datum — WGS 84, EPSG:4326, measured; mean sea level, assumed rather than modelled, so not EGM96 or EGM2008.
Coverage — global, ocean and land, with a companion grid naming each cell's source type.
Licence — public domain. GEBCO states the Grid is placed in the public domain and may be used free of charge, and separately asks that you acknowledge the source — a request beside a dedication, not a condition.
How to get it — the products page is the portal, current release GEBCO_2026; the CEDA directory holds the 2024 GeoTIFFs, of which one quadrant is 933,255,948 bytes.
Verdict — open, a download in practice: 890 MiB, one raster row per strip, no overviews. A third-party land COG at gebco2023.s3.valeria.science is 4,617,609,479 bytes, with a wrong content-type from its host.

USGS 3DEP — 10 m, United States only

If your area of interest is in the United States, stop reading the global list — this is an order of magnitude better.

Resolution — 1/3 arc-second, about 10 m, for the seamless product; 1 m, 1/9 arc-second and 5 m products exist.
CRS and vertical datum — NAD83, EPSG:4269, measured and confirmed in the metadata sidecar; NAVD 88 in metres per that sidecar, absent from the GeoTIFF header, so software reading only the file cannot warn you.
Coverage — the 48 conterminous states, Alaska, Hawaii and US territories.
Licence — public domain: the metadata sidecar records access constraints as "None" and asks only for acknowledgement.
How to get it — one staged tile covering San Francisco, 222,936,410 bytes, a true COG at 10812 by 10812, five overviews.
Verdict — open, with a cross-origin header, and still download it and drop the file in: 212.6 MiB is far past what a browser viewer should pull. The USGS programme page returned 403 to our script under two user agents and 200 to a full Chrome one, so that refusal is about the probe.

AWS Terrain Tiles — the whole world, 123 KiB at a time

A mosaic of elevation sources pre-cut into a web tile pyramid: not the dataset you analyse, the one you use when you want terrain on a map now.

Resolution — zoom-dependent, so a single number is meaningless; the tile we probed is zoom 11 at 512 by 512.
CRS — EPSG:3857. Its header reports a pixel size of 38.2, and those are 3857 units, not metres — read as ground resolution without the latitude factor, that is a wrong number.
Vertical datum — there isn't one. It mosaics SRTM, NED and others, so none applies.
Coverage — global, no internal overviews: the tile scheme is the pyramid.
Licence — per contributing source: the registry points at the Tilezen attribution page, listing each dataset's terms. Never call this public domain.
How to get it — the registry entry is the portal; one zoom-11 tile is a file, only 125,883 bytes.
Verdict — open, readable by a browser application, the smallest real raster in this run.

Brazil and Latin America

Brazil: the national portal is gone, and the answer is a global tile

This is the part that changed most since the Portuguese-language tutorials were written.

TOPODATA, the INPE project that was the canonical Brazilian DEM, is unreachable. Its hostname dsr.inpe.br does not resolve in DNS, over either protocol, and it is the whole host rather than one path — and an archive snapshot is a citation of a dead page, not a source. So the practical answer is a global source above, and we probed the São Paulo tile of each: Copernicus GLO-30 for a surface model, NASADEM for bare earth, SRTM for continuity.

For supporting layers, three URLs at one institution gave three answers — the whole argument for naming URLs rather than institutions.

IBGE, three URLs, three answers — the downloads portal 403s behind a Cloudflare challenge, to a script and a full browser user agent alike, so it is not linked here; geoftp.ibge.gov.br answers 200 with an open cross-origin header; and the boundary API answers 200, the easiest way to get Brazilian administrative outlines as GeoJSON.
INDE and Embrapa — geoservicos.inde.gov.br answers 200 but catalogues services rather than files, and Embrapa's "Brasil em Relevo" answers 206, but the URL is a download page rather than a file, so it is not linked.

Argentina, Chile, Colombia and Mexico

Portals rather than open buckets, so treat them as places to start a session. None published a licence or vertical datum we could cite today.

Argentina — the IGN publishes MDE-Ar v2.1 at a stated 30 m resolution with vertical accuracy around 2 m, plus 5 m and 0.5 m products. Start at the IGN elevation page; download from the geoportal.
Chile — the national geoportal is alive; the address most lists print takes two redirects to reach it, so that link is the final one.
Colombia — Colombia en Mapas, the IGAC front door.
Mexico — INEGI's Continuo de Elevaciones Mexicano offers 15 to 120 m grids, plus 5 m and 1.5 m LiDAR.

Test files you can use

Probed on 11 September 2026, and every "paste" line was also loaded in a real headless Chrome against the live viewer that day — a header check alone cannot promise a browser can read a file. One rule governs the list: the viewer's URL box downloads the entire file before rendering anything, so a pasted URL has to be small and everything else is "download it and drop the file in".

Two of the three come from OpenTopography, whose S3 bucket is the only host we measured that both allows cross-origin requests and exposes Content-Range to page JavaScript. Its download API is a different thing, and wants a key.

Paste: SRTM GL1 over São Paulo — . In the GeoTIFF viewer today: valid COG, 3601 by 3601, one Int16 band, EPSG:4326, 80 tiles in four seconds, 16,520,177 bytes over eleven requests — one whole-file 200 of 15,340,529 bytes and ten range reads costing 1,179,648 more. Licence: NASA data, cite the SRTMGL1 DOI.
Paste: NASADEM bare earth over São Paulo — , same dimensions and CRS, also a valid COG. Load both to compare bare earth against the original radar. Licence: NASA data, cite the NASADEM DOI.
Paste: one AWS Terrain Tile, to prove your setup works — , 125,883 bytes: tiled, no overviews, 512 by 512, Int16, EPSG:3857, on screen in about a second. Licence: per contributing source.
Download and drop: Copernicus GLO-30 — pasted into the viewer, the São Paulo tile fails at the CORS preflight with zero bytes transferred. Licence: the Copernicus terms in the bucket readme.
Download and drop: USGS 3DEP — every header on the staged tile says yes; 212.6 MiB through a URL box says no. Licence: public domain.
Download and drop: GEBCO — the 2024 quadrant, or the third-party land COG at 4.3 GiB. The quadrant is one row per strip with no overviews, so a viewer flags it striped and large. Licence: public domain, citation requested.

Famous and dead: stop looking for these

You have probably already tried some of these, which is why they are here.

TOPODATA at dsr.inpe.br — DNS does not resolve, over either protocol. The most-cited Brazilian elevation source on the Portuguese-language web, and gone.
The LP DAAC data directories at e4ftl01.cr.usgs.gov — the SRTMGL1 and NASADEM paths older tutorials print both return 404. The server is alive, so you get a tidy Not Found rather than a timeout, which reads like a typo rather than a move.
The LP DAAC product pages at lpdaac.usgs.gov — not dead, but every one now 301s to the earthdata catalogue URLs above. A list still printing that form is stale.
GEBCO's old data_and_products path — a 301 to the products page we link. Same story.
JAXA's own AW3D30 product-description PDF — the link on JAXA's live dataset page redirects to a URL-change notice. The dataset is fine; its documentation link is not, which is why the vertical datum entry above reads as it does.

One correction the other way, because calling a live source dead is the worse mistake: the USGS 3DEP programme page is not blocked. It refused our probe and serves a real browser fine.

What to do with the file once you have it

Read the header before the pixels. gdalinfo on the file — or on a /vsicurl/ URL, which reads only the header over range requests — gives the CRS, pixel size, block layout and whether overviews exist. That is how every measured claim here was made.

Then check the layout if you intend to serve it. A DEM internally tiled with overviews is a COG and a web client can read a window of it; one stored row by row cannot, at any size. The COG checker answers that without uploading anything, and our COG guide covers conversion.

Finally, write the vertical datum down somewhere your future self will find it: 5773 for the SRTM family, 3855 for Copernicus, NAVD 88 for 3DEP. A DEM whose vertical reference you have lost is a raster of plausible-looking numbers. For basemaps under it, we keep a list of free WMS and WMTS services on the same principle.

Managing rasters like these across a team — converting, tiling, storing and serving them so field staff can open them on a phone — is where this stops being a command and becomes a pipeline. That is Geodocs.

Free DEM and Elevation Data Sources, Verified in September 2026