BRENNEN

Industries / Transport and vehicles / Aerospace and defense / Space and satellites

Earth observation

Earth observation is taking pictures and other measurements of the land, sea and air from satellites, and selling or giving away the data and the maps made from it.

7

Milestones achieved, from 1960 to 2025.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1960

    First weather satellite

    NASA launched TIROS-1 on 1 April 1960 to test whether television pictures of cloud cover taken from orbit could be put to use. NASA's spacecraft catalog calls it the first weather satellite. Its two small TV cameras sent pictures straight to a ground station, or stored them on tape when the station was out of range, until a power failure on 15 June 1960 ended useful pictures.

    nssdc.gsfc.nasa.gov

    Credit

    NASA

  2. 1999

    Satellite pictures for sale

    On 12 October 1999 the company Space Imaging released what it called the "world's first high-resolution commercial satellite image of the Earth," a black-and-white picture of Washington, D.C., taken by its IKONOS satellite. Resolution here means the size of the patch of ground one pixel covers; IKONOS gave 1 m in black and white and 4 m in color.

    irp.fas.org

    Credit

    Space Imaging

  3. 2008

    Landsat archive made free

    Landsat is the US government's long-running series of land-imaging satellites. Before 2008 a Landsat scene cost from $400 to as much as $4,000, and the best year for sales, fiscal 2001, saw 25,000 scenes delivered, by the US Geological Survey's (USGS) count. Barb Ryan of the USGS and Michael Freilich of NASA signed a policy in January 2008 that made the archive free. More than 1 million scenes were downloaded in fiscal 2009, and more than 160 million by 2023; a 2017 USGS study put the yearly economic benefit at $3.45 billion worldwide.

    LessonTaking the price off public data raised its use about forty times within a year.

    usgs.gov

    Credit

    USGS and NASA

  4. 2021

    Near-daily pictures, measured

    PlanetScope is Planet Labs' fleet of small satellites that take pictures with 3 m pixels. A study led by David Roy at Michigan State University, with one Planet co-author, counted every public PlanetScope image from December 2019 to November 2020: 175.8 million images from 100 to 133 working satellites a month. Half of all land points were seen on average every 30.3 hours or more often, but only 9.6% averaged under 24 hours between pictures, and 9% of land points, mostly inside Greenland and Antarctica, had no pictures at all.

    doi.org

    Credit

    Planet Labs (fleet); David Roy, Haiyan Huang, Rasmus Houborg and Vitor Souza Martins (study)

  5. 2022

    A radar twin is lost

    Sentinel-1 is the radar mission of the European Union's Copernicus program; radar sees through cloud and darkness. With two satellites, 1A and 1B, it could image a place as often as every six days. On 23 December 2021 a fault in the power supply of Sentinel-1B's radar left it unable to deliver data, and on 3 August 2022 ESA and the European Commission declared its mission over. ESA re-planned Sentinel-1A and used radar data from other missions, such as Radarsat-2, TerraSAR-X, COSMO-SkyMed and PAZ, to cover some gaps.

    LessonA pair of satellites with no spare ready loses half its coverage until a replacement is built and launched.

    esa.int

    Credit

    ESA and the European Commission

  6. 2024

    Foundation models get tested

    A geospatial foundation model is an AI model trained on large amounts of unlabeled satellite imagery, meant to be adapted cheaply to many mapping jobs. In December 2024 Valerio Marsocci and 14 co-authors published PANGAEA, one standard test across varied datasets, tasks, sensors and resolutions, and found that the popular open models "do not consistently outperform" ordinary supervised models such as UNet, a standard network trained for one task only.

    LessonA model trained on more data is not better until it is tested on the same terms as the simpler method it would replace.

    arxiv.org

    Credit

    Valerio Marsocci and co-authors

  7. 2025

    Ten-meter Earth summaries

    In July 2025 Google DeepMind released AlphaEarth Foundations, an AI model that condenses many satellite and other data sources into an embedding, a short list of numbers, for every 10 m square of land and coastal water, published yearly in Google Earth Engine. Google says it had a 24% lower error rate than the other models it tested, a figure from its own tests. Outside groups now use it routinely; arXiv listed 65 papers mentioning it on 18 November 2026. In one outside test, published in September 2026, Philipp Stark, Alexandros Sopasakis and Ola Hall found that for mapping burned land the embeddings beat simpler summaries of Sentinel-2 images by 0.05 to 0.13 in ROC AUC, an accuracy score that runs from 0.5 for guessing to 1 for perfect, inside one fire. On a fire in another region the advantage disappeared, until eight labeled blocks from that region were added.

    deepmind.googlearxiv.org

    Credit

    Google DeepMind

Watch for

  1. Free radar of all land

    NISAR, a radar satellite built by NASA and the Indian Space Research Organisation (ISRO), launched on 30 July 2025. NASA's Jet Propulsion Laboratory (JPL) says it has been imaging nearly all of the planet's land and ice twice every 12 days, producing dozens of terabytes of data a day, and the US team began releasing calibrated L-band radar products on 20 July 2026. NASA's data notices label many of those products "Provisional Version 1" and record a permanent data gap from 27 July to 10 August 2026.

    Achieved whenthe first full year of data is out as final calibrated products and outside users publish measurements checked against ground data.

    jpl.nasa.govearthdata.nasa.govearthdata.nasa.gov

    Claimed by

    NASA and ISRO, July 2026

  2. Satellites that pick shots

    Clouds often hide the ground from camera satellites. From July to September 2025 JPL, with the Irish company Ubotica and the British company Open Cosmos, flew "dynamic targeting" on the small satellite CogniSAT-6. The satellite tilts forward to look 60 to 90 seconds ahead along its path, checks that view for cloud on an onboard Intel Myriad processor, and uses the result to aim its main camera on the same pass.

    Achieved whenan operator uses it in routine service and publishes, or an outside party measures, the share of usable cloud-free pictures with and without it.

    ai.jpl.nasa.gov

    Claimed by

    JPL, Ubotica and Open Cosmos, 2025

Open

  1. 01

    Date-check a satellite picture

    A free tool that takes a satellite picture posted online, with its claimed place and date, and reports whether the claim fits. It pulls the free Sentinel-2 and Landsat images of that place from the days around the claimed date, compares what is on the ground, and checks the direction of shadows against where the sun would have been.

    The brief
    Credit · unclaimed
  2. 02

    Check sellers' map accuracy

    A public table of how far features in the free sample images of several commercial satellite companies sit from their true positions on the ground, measured the same way for each company against independent survey-grade references.

    The brief
    Credit · unclaimed

Open 01 · the brief

Date-check a satellite picture

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A free tool that takes a satellite picture posted online, with its claimed place and date, and reports whether the claim fits. It pulls the free Sentinel-2 and Landsat images of that place from the days around the claimed date, compares what is on the ground, and checks the direction of shadows against where the sun would have been.

Why it is still open

A 2023 verification guide for journalists still lists these as manual steps: check the original provider, compare with other sources, test the shadows with the website suncalc.org, and consult an expert. A 2025 method from the USGS and partners, called Sundial, measures shadow directions to estimate the time of day a picture was taken, within 2.1 ± 3.4 minutes on 16 WorldView satellite images and six aerial images, but it assumes the date and place are known and does not test a claim. Bellingcat's free ShadowFinder, released in 2024, works the other way: it takes an object's height, its shadow's length and a date and time as given, and narrows down where on Earth the picture could have been taken. A search found no tool that checks a claimed date and place against the free archives.

institute.aljazeera.netusgs.govbellingcat.com

Built with what exists

Sentinel-2 images from the Copernicus Data Space Ecosystem or Microsoft Planetary Computer; Landsat images from the USGS; Python with Skyfield or pvlib for the sun's position; the Sundial method; the Maxar Open Data Program, a public catalog of dated disaster images under the CC BY-NC 4.0 license, as test material.

registry.opendata.aws

  1. First step

    Take 20 dated Maxar Open Data scenes, measure the shadow direction by hand in five of them, and compare it with the sun direction computed for the stated date and time.

  2. Done looks like

    On 200 dated pictures, each offered once with its true date and once with a false date 3 to 60 days away, the tool marks at least 90% of the false claims as not fitting and no more than 5% of the true ones, with code and test set published.

Open 02 · the brief

Check sellers' map accuracy

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public table of how far features in the free sample images of several commercial satellite companies sit from their true positions on the ground, measured the same way for each company against independent survey-grade references.

Why it is still open

Companies now publish open sample archives. Umbra's radar archive holds more than 250 images worldwide under the CC BY 4.0 license, and Capella Space adds radar data to its open archive every quarter. Space agencies check companies one at a time, on the data each company supplies them. ESA's Earthnet Data Assessment Project (EDAP) assesses candidate missions for ESA's own program, and its page lists no table across companies. In May 2026 NASA's Commercial Satellite Data Acquisition program reported that Umbra's image sharpness matched the company's specifications but that its geolocation accuracy "did not universally align" with them. A search found such single-company reports, but no check that measures several companies' public images by one method and can be rerun when their processing changes.

registry.opendata.awsregistry.opendata.awsearth.esa.intscience.nasa.gov

Built with what exists

The Umbra and Capella open radar archives and the Maxar Open Data Program on Amazon Web Services; national lidar and aerial photo maps, such as the USGS 3D Elevation Program, as the ground reference; Python with GDAL and rasterio.

  1. First step

    Pick 10 Umbra open images of US cities, mark 10 clearly visible fixed points in each, and measure their offsets from USGS aerial photo maps.

  2. Done looks like

    For at least three companies and 30 images each, the median and 90th-percentile position error in meters, measured on at least 10 independent ground points per image, set beside each company's stated accuracy, with the points and code published.