Global Landslide Segmentation

Description

Code release for the paper "Automating global landslide detection with heterogeneous ensemble deep-learning classification"

Ganerød, A. J., Franch, G., Lindsay, E., & Calovi, M. (2023). Automating global landslide detection with heterogeneous ensemble deep-learning classification. arXiv preprint arXiv:2310.05959.

preprint: https://arxiv.org/abs/2310.05959

data, code, and models: https://zenodo.org/records/13471162

How to run

Install dependencies

# clone project repository
git clone https://github.com/franchg/global-landslide-mapping
cd global-landslide-mapping

# create and activate environment
# adjust python version if needed
PYTHON_VERSION=3.11
python$PYTHON_VERSION -m venv .venv
source .venv/bin/activate

# install requirements
pip install -r requirements.txt

Download data from Zenodo by executing the script in the data/ folder

cd data
bash download_data.sh

Download models from Zenodo by executing the script in the models/ folder

cd models
bash download_pretrained_models.sh

See the notebook notebooks/test_enseble_classification.ipynb for running the models on the test data and generating the results.

Train from scratch

Train an ensemble of models with one of the following configurations contained in the folder configs/experiment/:

• all_channels - all channels
• sentinel1_2 - Sentinel-1 and Sentinel-2 bands
• sentinel1 - Sentinel-1 bands
• sentinel2 - Sentinel-2 bands

# train on GPU using all channels as input
# this will train an ensemble of 90 models (2 lr x 5 losses x 9 models)
# the result will be saved in the folder `logs/train/multirun/`
python src/train.py --multirun hparams_search=grid_search trainer=gpu experiment=all_channels.yaml