Awesome

Landslide mapping from Sentinel-2 imagery through change detection

Code repository for the paper "Landslide mapping from Sentinel-2 imagery through change detection" (accepted at IGARSS 2024)

Get started

To run code in this repository, you first have to install required libraries

pip install -r requirements.txt

Then, you have to download the S2 image dataset for each inventory by running the Jupyter Notebooks in processing/.

To train a model, you can run train.py specifying hyperparameters as arguments. Arguments can be given either as command line arguments (e.g. python train.py --encoder="resnet50") or specified in a YAML configuration file (e.g. python train.py --yaml="configs/bbunet.yaml"), or even both. If both a YAML config and command line arguments are specified in the command, command line arguments eventually override YAML arguments. If an argument is not specified, a default value is used, as specified in src/arg_parser.py).

To specify a pretrained backbone through the --pretrained_encoder_weights argument, you have to manually download the corresponding "full checkpoint" from the SSL4EO-S12 repository first.

In configs/ you can find hyperparameters used to train the models reported in the paper. Thus, the experiments in the paper can be run simply by:

python train.py --yaml="configs/model.yaml"

Checkpoints

The models trained in the paper can be found at this link. The following code is used to load a model:

import torch
from argparse import Namespace
from src.model_utils import make_model

# load trained checkpoint
ckpt_path = "path/to/model.pt"
ckpt = torch.load(ckpt_path, map_location='cpu')
hparams = ckpt['hparams']

# load model architecture
model = make_model(hparams)

# load trained weights into model
model.load_state_dict(ckpt['state_dict'])