bayes-forest-structure

Code accompanying our paper "Country-wide Retrieval of Forest Structure From Optical and SAR Satellite Imagery With Bayesian Deep Learning" [link to paper]

Code organization

• ./config configuration files for data preprocessing, training and testing
• ./src main project code
  • ./src/scripts helper scripts for preprocessing, evaluation etc.

Getting started

Prerequisites

• python >= 3.8
• numpy >= 1.19
• pytorch >= 1.9
• scikit-image >= 0.18
• rasterio >= 1.1.5
• gdal >= 3.0
• fiona >= 1.8.13
• latextable

Inference

To apply our model to your data, preprocess the SAR data using the /src/scripts/preprocess_s1.py script and then reproject it to the Sentinel-2 tile. Then use /src/scripts/predict_tile.py to process the tile, passing the locations to the optical and SAR files as arguments. Make sure to also pass the paths of the config.yaml file (containing network configurations and statistics for data normalization) and of all five network checkpoints via the respective command line args. The files can be downloaded from here. The output tiles will be written to the location passed as --out_dir.

Training

For improved efficiency during training, the preprocessed and reprojected Sentinel images as well as the ground truth should be pickled first using the /src/scripts/prepare_dataset.py script. For training the /src/train.py script (along with the /config/resnext.yaml config file) is used.

Citation

@article{becker2023country,
  title={Country-wide retrieval of forest structure from optical and SAR satellite imagery with deep ensembles},
  author={Becker, Alexander and Russo, Stefania and Puliti, Stefano and Lang, Nico and Schindler, Konrad and Wegner, Jan Dirk},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={195},
  pages={269--286},
  year={2023},
  publisher={Elsevier}
}