Geospatial ML2024Shipped

Glacier U-Net

Glacier boundaries segmented from 5-band Sentinel-2 imagery.

A U-Net for binary semantic segmentation of glacier boundaries on 5-band Sentinel-2 imagery, with a custom geospatial loader that tiles multispectral GeoTIFFs and normalizes each band against training-set statistics.

Source Codearrow_outward Next: Cadencearrow_forward

livelast push Dec 21

~/glacier

Project Telemetry

Dice

0.87

IoU

0.78

Training patches

1,200

Spectral bands

How it works

~/glacier/architecture · step 1/2

5-band GeoTIFFs

tiling + band norm

U-Net

mask stitching

Dice / IoU on held-out regions

01Reported Dice and IoU on spatially disjoint held-out regions. Pixel accuracy is uninformative at ≈12% glacier pixels.

02Compared BCE, Dice and combined weighted losses, and analysed operating points via ROC.

Project Info

DomainMachine Learning

RoleAuthor

Year2024

StatusShipped

Tech Stack

PythonPyTorchU-NetSentinel-2GeoTIFF