Geocodis Builtup Areas

Machine Learning based algorithm for Built up area detection.


Builtup Areas Sentinel-2 uses a U-Net convolutional deep neural network to detect built-up areas. The image is first segmented into smaller chunks which are than classified by a trained model and finally reassembled into the final image. Relevant packages: keras, tensorflow, rasterio, scikit-image, scipy.

Supported Workflows

  • Sentinel-2 L2A Analytic (GeoTIFF) -> Builtup Areas Sentinel-2

Algorithm Training Details

The model was trained on data from Uganda which may limit its performance in other areas. The model sometimes detects dirt as built up areas and sometimes fails to detect larger structures and small huts. We plan to improve detection accuracy in next releases.

Algorithm accuracy: average more than 80%

Input data format

Results from block Sentinel-2 L2A Analytic (GeoTIFF)

Parameter: probabilities - if false (default) the result is a binary classification indicating where the built-up areas are. If set to true, the resulting image represents the likelihood of built-up areas as estimated by the algorithm.

Output data format

  1. If ‘probabilities’ parameter is false: GeoTIFF (byte) with binary classification (0, 1), dtype : uint8, resolution: propagated (10m x 10m)
  2. If ‘probabilities’ parameter is true: GeoTIFF (float) with values between 0 and 1,
    dtype : float, resolution: propagated (10m x 10m)

Terms & Conditions

View the End User License Agreement conditions.

Ready to get started?

Get in touch or become a partner.

Contact salesBecome a partner