AI · very high resolution imagery

Detecting buildings, vehicles and boats in very high resolution imagery

A YOLO detector trained on xView (30 cm images, 60 classes), applied to another very high resolution satellite image. Detections are grouped into 14 classes, six of which are outlined with SAM.

Expertise France · 2026

Object detection application: settings panel and map of small boats detected in a marina
The application: live detection or loading of results, grouping into 14 classes, SAM outlines; on the map, filters by score and class. Here, a marina.

Who it is for

For counts and inventories over large areas (buildings, vehicles, boats), to be checked before any official use.

What the application does

Detection over an area

Rectangle drawn on the image, up to 10,000 x 10,000 px, detection on GPU and grouping of the 60 classes into 14.

SAM outlines

SAM 2.1 outlines buildings, ships, small boats, storage tanks, aircraft and helicopters, with their area.

Filters and export

Minimum score, displayed classes and counts without re-running. GeoPackage and CSV export.

How it works

  1. 01

    Data

    xView: 30 cm WorldView-3 images annotated in 60 classes, cut into 640 px tiles. Validation on 10% of the scenes, as the official validation labels are not public.

  2. 02

    Training

    YOLO (Ultralytics) on GPU, saved at every epoch and resumed after an interruption.

  3. 03

    Detection

    640 px tiles with 20% overlap, merged detections, then SAM outlines from the boxes.

Screenshots

Small boats detected in a marina, with boxes and outlines
Marina: "small boat" class, YOLO boxes and SAM outlines. Detections computed on a 30 cm image, shown over the Esri basemap.
Buildings and vehicles detected in a residential area
Residential area: 212 buildings and 20 small vehicles at score ≥ 0.10, a low threshold that also keeps false detections.

Known limitations

  • Imbalanced classes: buildings and small cars dominate, rare classes are detected less well.
  • Model trained on WorldView-3 and applied here to another 30 cm image, with no performance measured on it.
  • Axis-aligned boxes, poorly suited to long oblique objects.

Going further

  • Oriented boxes for ships and aircraft
  • Fine-tuning on local imagery (drone, Pléiades Neo) with a few annotations

Role

  • Data preparation
  • GPU model training
  • Demonstration application

Technologies

PyTorch · YOLO (Ultralytics) · SAM 2.1 · xView · GeoPackage · ipyleaflet · Voilà · GPU

Context

Demonstrator designed and built in 2026 by Guillaume Rieu, author of the Earth Innovation Labs specifications, as a volunteer contribution to the acceptance testing of these platforms in Kenya for Expertise France. It runs in the platform's JupyterHub environment.

A similar need?

These demonstrators can serve as the basis for a tool adapted to your data, your area or your question.

Contact us

Other case studies