Use case: across industries

Analysing satellite imagery with AI

AI turns satellite images into information: a model learns from labelled examples what an object, a land cover type or a change looks like, then finds it across large areas automatically. Typical tasks are counting objects, classifying land cover and flagging changes. Models need examples that match your region and sensor, and their results need checking.

Reference
UC-30
Technologies
  • Artificial intelligence
  • Earth observation
Industry
Challenge
Asset and infrastructure monitoring
Updated
Published by
LETO SPACE

01Technology

How it works

A satellite image is a grid of measured values. AI models learn from labelled examples which patterns in that grid stand for a building, a ship, a field or a clearing.

Satellites now give a consistent, global view of forests, oceans and growing cities, but the images have little value without processing chains that turn pixel values into information for decision makers. Machine learning, especially deep learning, is increasingly part of these chains Source: IEEE Geoscience and Remote Sensing Magazine, 9 September 2024IEEE Geoscience and Remote Sensing MagazineArtificial Intelligence to Advance Earth Observation: A review of models, recent trends, and pathways forward (Tuia, Schindler, Demir, Zhu et al.)9 September 2024Open source (opens doi.org).

The models solve a few recurring tasks:

Object detection and counting

Because of the large volume of images collected every day, automatic detection and segmentation of objects are a constant need in Earth observation Source: Remote Sensing (MDPI), 12 January 2024Remote Sensing (MDPI)Remote Sensing Object Detection in the Deep Learning Era: A Review (Gui, Song, Qin and Tang, Remote Sensing 16, 327)12 January 2024Open source (opens doi.org). One example: the DeepSolar framework used deep learning on satellite imagery to locate solar installations and estimate their size across the contiguous United States Source: Joule, 19 December 2018JouleDeepSolar: A Machine Learning Framework to Efficiently Construct a Solar Deployment Database in the United States (Yu et al., Joule 2, 2605 to 2617)19 December 2018Open source (opens doi.org). Its use for solar power is described under rooftop solar potential.

Classification

The model assigns each pixel or area to a class, such as forest, water, cropland or sealed ground.

Change detection

The model compares images of different dates and flags what has changed, for example new buildings or cleared forest.

Foundation models

These models are pre-trained on large data sets with self-supervised learning and then fine-tuned for a specific task. One example is Prithvi, released openly by NASA and IBM in 2023 and trained on Harmonized Landsat and Sentinel-2 imagery; it has been applied to flood mapping and to the detection of burn scars Source: NASA Earthdata, 3 August 2023NASA EarthdataIBM Research and NASA IMPACT release open-source geospatial foundation model on Hugging Face3 August 2023Open source (opens earthdata.nasa.gov).

Three tasks on the same scene: detection finds and counts objects, classification assigns each area to a class, and change detection compares two dates. A foundation model is pre-trained on large data sets, fine-tuned with examples of your task and then tested on your region. Schematic scene.

AI also runs on board satellites. ESA's Φsat-1, launched in 2020, was the first experiment to demonstrate AI for Earth observation in this way: it filtered out unusable images in orbit, so that only usable data was sent to Earth Source: European Space Agency (ESA), accessed 8 October 2026European Space Agency (ESA)Φsat (mission page)accessed 8 October 2026Open source (opens esa.int).

02Data

Which data and tasks

TaskQuestion it answersTypical imagery
Object detection and countingWhere are the objects, and how many are there?Very high resolution optical images, for example Pléiades Neo at 30 cm
ClassificationWhat covers the ground in each area?Optical images such as Sentinel-2 at 10 m
Change detectionWhat has changed between two dates?Time series from the same sensor
Fine-tuning a foundation modelCan a pre-trained model be adapted to my task?Imagery similar to the model's training data

Which objects a model can find depends on the pixel size. Sentinel-2 records its finest bands at 10 m Source: European Space Agency (ESA), SentiWiki, accessed 6 October 2026European Space Agency (ESA), SentiWikiSentinel-2 missionaccessed 6 October 2026Open source (opens sentiwiki.copernicus.eu), enough for fields and forest stands but not for cars or small structures. Pléiades Neo records panchromatic images at 30 cm Source: European Space Agency (ESA), Earth Online, accessed 6 October 2026European Space Agency (ESA), Earth OnlinePléiades Neoaccessed 6 October 2026Open source (opens earth.esa.int).

03Deliverables

What you get

04Limits

Limits

  • Limit 01

    Training data must fit

    Objects look different from region to region and in images from different sensors, optical or radar, and platforms. Missing or scarce labels are a known limit of deep learning in Earth observation Source: Remote Sensing (MDPI), 12 January 2024Remote Sensing (MDPI)Remote Sensing Object Detection in the Deep Learning Era: A Review (Gui, Song, Qin and Tang, Remote Sensing 16, 327)12 January 2024Open source (opens doi.org). A model trained elsewhere needs testing on your region and your imagery.

  • Limit 02

    Pixel size

    A model cannot find what the imagery does not resolve. Check the pixel size against the smallest object you need to see before you choose the imagery and the model.

  • Limit 03

    Results need checking

    Every model makes errors. Before a result feeds a decision, compare it with reference data you trust, and keep checking samples once the model is in use.

05Next steps

How we help

Whether a ready service, a platform or a custom model fits depends on your question. In a strategy engagement we work out which question the model must answer, how accurate it must be and which imagery it needs, and we then write the requirements for choosing a provider. Spacegraph, our discovery platform, has marketplace categories for object detection and counting and for change detection.

For teams that want to build skills in-house, the LETO Academy offers the course "GeoAI: Deep Learning for Earth Observation" (opens academy.leto.space), at intermediate level and by invitation, with exercises in Python.

LETO Academy

Our self-paced courses on AI, digital skills and space-enabled solutions, in English and German, some open, some by invitation.

Courses in EN and DEBrowse Academy courses (opens academy.leto.space)

Skills and workshops

06Use cases

Frequently asked questions

Q1What is GeoAI?

GeoAI is the use of machine learning on geographic data such as satellite and aerial images. Satellite data has little value without processing chains that turn pixel values into information for decision makers, and machine learning, especially deep learning, is increasingly part of these chains Source: IEEE Geoscience and Remote Sensing Magazine, 9 September 2024IEEE Geoscience and Remote Sensing MagazineArtificial Intelligence to Advance Earth Observation: A review of models, recent trends, and pathways forward (Tuia, Schindler, Demir, Zhu et al.)9 September 2024Open source (opens doi.org).

Q2Do we need to train our own model?

Not always. Foundation models such as Prithvi, released openly by NASA and IBM in 2023, are pre-trained on large data sets and can then be fine-tuned for a specific task Source: NASA Earthdata, 3 August 2023NASA EarthdataIBM Research and NASA IMPACT release open-source geospatial foundation model on Hugging Face3 August 2023Open source (opens earthdata.nasa.gov). Fine-tuning still needs examples from your task, and the result needs a test on your region.

Q3How do we start a pilot with AI on satellite imagery?

Define one question the model must answer, for example how many objects of a type stand on a set of sites, and collect a set of examples you trust to check the results. Then compare an existing service with a model adapted to your data.

Sources (7)

  1. IEEE Geoscience and Remote Sensing Magazine, 9 September 2024Artificial Intelligence to Advance Earth Observation: A review of models, recent trends, and pathways forward (Tuia, Schindler, Demir, Zhu et al.) (opens doi.org)
  2. Remote Sensing (MDPI), 12 January 2024Remote Sensing Object Detection in the Deep Learning Era: A Review (Gui, Song, Qin and Tang, Remote Sensing 16, 327) (opens doi.org)
  3. Joule, 19 December 2018DeepSolar: A Machine Learning Framework to Efficiently Construct a Solar Deployment Database in the United States (Yu et al., Joule 2, 2605 to 2617) (opens doi.org)
  4. NASA Earthdata, 3 August 2023IBM Research and NASA IMPACT release open-source geospatial foundation model on Hugging Face (opens earthdata.nasa.gov)
  5. European Space Agency (ESA), accessed 8 October 2026Φsat (mission page) (opens esa.int)
  6. European Space Agency (ESA), SentiWiki, accessed 6 October 2026Sentinel-2 mission (opens sentiwiki.copernicus.eu)
  7. European Space Agency (ESA), Earth Online, accessed 6 October 2026Pléiades Neo (opens earth.esa.int)

Contact

Detect, classify or compare dates: tell us what your imagery should answer.

Prefer to talk?

Tell one of our experts about your project. Together we work out where you stand, whether satellite data and AI can help and what a sensible next step would be.

Both calls are free of charge and take place online. You choose a date in our booking calendar.