Model lab · the paper's demo

Classify every pixel of your satellite image

The demo described in Section 3.6 of Khadka & Pradhan (2026): upload an image and the U-Net segments it into the seven DeepGlobe land-cover classes. Choose the original DeepGlobe model or the Nepal-adapted one, feed the image exactly as the paper did or at its true scale, and switch post-processing on or off.

Satellite or aerial image (JPG or PNG, taken from directly above)

The image never leaves your device: the model runs in your browser.

How the image is fed to the network As in the paper: resize the whole image to 224 × 224 True scale: resample to 5.46 m per pixel

Model Post-processing: remove regions smaller than 64 pixels

Tips

The original model was trained on 50 cm DeepGlobe imagery, so it works best on sub-metre images such as Google Earth or drone mosaics taken straight down. For Sentinel-2 or Landsat images use the Nepal-adapted model at true scale (10 m or 30 m per pixel). No image to hand? Map any place in Nepal directly from satellite imagery.