terrain-referenced-3d-glacier-training-code

Prithvi-SDT Training Code

This repository contains only the training code for the final Prithvi-SDT model used in the manuscript.

Files

• glacier_dataset_hybrid.py: loads the glacier training dataset, including multi-temporal optical image stacks, DEM-derived terrain layers, segmentation masks, and RGI-region test splits.
• model_prithvi_sdt.py: defines the final Prithvi-SDT architecture, combining a Prithvi Earth-observation backbone, a detail branch, and terrain-detail fusion for glacier-area segmentation.
• train_prithvi_sdt.py: trains, resumes, validates, checkpoints, logs, and evaluates the full Prithvi-SDT model.

Prithvi Backbone

Prithvi-SDT uses the IBM/NASA Prithvi-EO-2.0 300M model as the Earth-observation backbone through TerraTorch. The Prithvi backbone source code and pretrained weights are not duplicated in this repository. Before training, download the official Prithvi-EO-2.0 300M checkpoint from:

https://huggingface.co/ibm-nasa-geospatial/Prithvi-EO-2.0-300M

Then set --weight-path to the local checkpoint path, for example:

python train_prithvi_sdt.py \
  --dataset-root /your/path/to/glacier_dataset \
  --train-root /your/path/to/output_dir \
  --weight-path /your/path/to/Prithvi_EO_V2_300M.pt

All paths in the released code are placeholders such as /your/path/to/... and should be replaced with user-specific dataset, Prithvi checkpoint, and output locations before training.