SpaceNet Building Detection

This model is a high-performance U-Net architecture with residual connections, trained to detect precise building footprints from high-resolution satellite imagery. It was trained on the SpaceNet Area of Rio de Janeiro dataset. This demo showcases how the deep learning model can segment and highlight individual building structures in complex urban environments. More details can be found here.

The user needs to provide a 3-band (RGB) satellite geotiff image (.tif), or a standard image file (.png, .jpg). The model will output the raw imagery, a heatmap of the predicted building score, and the image overlaid with the predicted building masks.

Examples