Vadstena: AI-based Automatic Land Cover

By Martina Bekrová
September 17, 2018

Geospatially accurate land cover in the form of georeferenced GeoTIFFs, utilitizing a fixed 8-bit palette of 25 land-cover classes, now forms a standard part of Vadstena’s output, obtained automatically for any Vadstena-processed dataset.

​Vadstena reality-capture system converts the real world into its digital counterpart. In its current shape and form, it takes aerial imagery and turns it into an array of products, including photorealistic textured 3D polygonal meshes, digital terrain models and orthophotomaps.

While Vadstena is better known for being the technology behind Melowntech’s immersive, web-VR experiences, our more recent R&D goes beyond mere extraction of 3D geometries. Thus, as a new type of Vadstena output with a huge potential for automatic production of valuable geospatial data, we introduce machine-learning-based, high-detail ​land cover.

Using state-of-the-art artificial intelligence techniques, we classify each pixel of every input image into 25 classes such as roof, tree, vehicle, grass, and many more:

Land Cover Class Legend

  • Roof
  • Facade
  • Terrace
  • Tree
  • Shrub
  • Structure
  • Object
  • Solar Panel
  • Vehicle
  • Train
  • Boat
  • Airplane
  • Wall
  • Retaining Wall
  • Stairs
  • Bridge
  • Impervious
  • Dirt Road
  • Railway
  • Sports Field
  • Water
  • Agriculture
  • Grass
  • Sand
  • Rock

As a result, geospatially accurate land cover, in the form of georeferenced GeoTIFFs, utilitizing a fixed 8-bit palette of land-cover classes, now forms a standard part of Vadstena’s output, obtained automatically for any Vadstena-processed dataset.

Since Vadstena is able to correlate its redundant input (each location is visible in multiple photos), we can combine multiple predictions from our convolutional neural network (CNN) to obtain even more accurate land-cover classification.

To make the process as robust as possible, we are maintaining a growing database of hand-labeled training data created from a variety of images taken at diverse locations around the world, by different camera systems and at varying resolutions. The goal is to make sure we can accurately predict land cover for any type of urban and natural landscape – from a cozy Danish town to a bustling East African city.

See below how the automatic land cover performed in some of the datasets we recently processed with Vadstena.

Hilleroed, Denmark

GSD 5cm, high-altitude, high-overlap oblique-camera-system imagery*

* Source data courtesy of COWI.

St. Louis, Missouri, USA

GSD 8cm, high-altitude, mid-overlap oblique-camera-system imagery

Trenčín, Slovakia

GSD 5cm, low-altitude, high-overlap UAV imagery

Dar es Salaam, Tanzania

GSD 8cm, low-altitude, high-overlap UAV imagery

Olomouc, Czechia

GSD 10cm, high-altitude, high-overlap oblique-camera-system imagery


Note that neither of the datasets shown above was used for training the CNN.

Martina Bekrová, Machine Learning Specialist, Vadstena

Behind the pretty pictures, there is a real world. Every human can easily tell a house from a tree. Martina is delving deep into deep learning in order to pass these skills on to Vadstena 3D reality-capture system, to discover the meaning behind pixel values.