For the paper and presentation please check the project page.
conda env create -f environment.yml
conda activate glsegenv
cd gl_seg_dl
- Download the glacier outlines:
bash ./scripts/download_outlines.sh
- Download the glacier rasters (stored here: https://huggingface.co/datasets/dcodrut/glacier_mapping_alps):
bash ./scripts/download_data.sh
The archived rasters have ~3Gb for each year (inventory one & 2023). After extracting the NetCDF rasters, we will need 20Gb for each year.
- Cross-validation splits & patch sampling:
python main_data_prep.py
- Compute the mean & stddev of the training patches:
python main_compute_data_stats.py
- Train the five models:
bash scripts/train.sh
(by default it runs on a single GPU; check the bash script for running on multiple GPUs) - Apply model on each glacier both on the inventory images and the 2023 ones:
bash scripts/infer.sh
- Estimate the areas based on the predictions from the previous step:
bash scripts/eval.sh