/veg-height-map-public

[SRS 2023] Countrywide Vegetation Height Estimation with Sentinel-2 and Deep Learning

Primary LanguagePython

veg-height-map-public

[SRS 2023] Countrywide Vegetation Height Estimation with Sentinel-2 and Deep Learning

🌲 Data Availability 🌲

The generated vegetation height maps of Switzerland (including both the mean & max height) are public accessible please download here

  • resolution: 10m-resolution
  • temporal coverage: 2017, 2018, 2019, 2020
  • spatial coverage: the whole Switzerland

demo_map_2019


🌿 Requirements

  • Python 3.8.5
  • PyTorch: 1.7.1+cu110 (gcc/6.3.0, cudnn/8.0.5, cuda/11.0.3)
  • HDF5/1.10.1
  • GDAL/3.1.2

🌳 Preproccessing

  • generate image stats
python -m scripts.calculate_stats --img=True --preproconfig=configs/preprocess.yaml
  • preprocess images
python -m scripts.preprocess_img --preproconfig=configs/preprocess.yaml
  • normalise images/labels
python -m scripts.calculate_stats --preproconfig=configs/train_spyr.yaml
  • normalise DTM
python -m scripts.calculate_stats --preproconfig=configs/train_spyr.yaml --dtm True

🍀 Training

  • train with DTM
python -m scripts.train --config=configs/train_spyr.yaml
  • train without DTM
python -m scripts.train --config=configs/train_spyr_nd.yaml

🪴 Inference

  • predict tile TMS with the model with DTM
python -m scripts.predictSet --pred_config_path=configs/predict_dtm_TMS.yaml --train_config_path=configs/train_spyr.yaml
  • predict tile TMS with the model without DTM
python -m scripts.predictSet --pred_config_path=configs/predict_nodtm_TMS.yaml --train_config_path=configs/train_spyr_nd.yaml

🌵 Evaluation

python -m scripts.eval --config=configs/eval.yaml

🌱 Citation 🌱

@article{jiang2023accuracy,
  title={Accuracy and consistency of space-based vegetation height maps for forest dynamics in alpine terrain},
  author={Jiang, Yuchang and R{\"u}etschi, Marius and Garnot, Vivien Sainte Fare and Marty, Mauro and Schindler, Konrad and Ginzler, Christian and Wegner, Jan D},
  journal={Science of Remote Sensing},
  pages={100099},
  year={2023},
  publisher={Elsevier}
}