AI tool
Map the land cover of any place in Nepal
Pick a place, choose the imagery, and the U-Net from Khadka & Pradhan (2026) labels every 5.46 m pixel as urban, agriculture, rangeland, forest, water or barren land. It runs entirely in your browser, with no sign-up and no upload, and gives you the map and the area of each class.
Selected: Kathmandu core (Durbar Square → Thamel) (27.7090, 85.3110)
How it works
The imagery is resampled to 5.46 m per pixel, the scale the network was trained at (DeepGlobe's 2,448-pixel, 0.5 m tiles resized to 224 pixels), with the same bilinear method used in training. Each year's Sentinel-2 cloudless mosaic is colour-normalised first. Regions smaller than 64 pixels (0.19 ha) take the class of the nearest larger region, the paper's morphological clean-up.
The Nepal-adapted model agrees with the independent ESA WorldCover map on about 72% of pixels on districts it never saw in training. Treat single maps as estimates. For whole districts with confidence intervals see the district pages, and to compare two dates use change detection.