A continuous DEM model. Please refer to our paper "A continuous digital elevation representation model for DEM super-resolution" (in ISPRS Journal of Photogrammetry and Remote Sensing).
If you find our work useful in your research, please cite:
@article{YAO20241,
title = {A continuous digital elevation representation model for DEM super-resolution},
journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
volume = {208},
pages = {1-13},
year = {2024},
issn = {0924-2716},
doi = {https://doi.org/10.1016/j.isprsjprs.2024.01.001}
}
pip install -r requirements.txt
Our code deeply relys on the "hydra" package, for every experiments it need to modify the corresponding config file. We provid the debug template for config files.
python hydra_run.py --cfg job --resolve ${config_file}
-
Download the data: TFASR30 and TFASR30to10 datasets, Pyrenees and Tyrol datasets.
-
Split Pyrenees and Tyrol datasets:
python make_dataset.py --dirRawDataset ${data_path}/terrain/pyrenees
# It should generate the split files in "./temp" dir. Move and rename the dir if you want. Similar processing for Tyrol dataset.
-
Generate json file for datasets: Modify your specific changes in "generate_json_for_dataset.py" file.
-
Modify the config file in "configs/datasets/${dataset_name}.yaml"
-
Generate the training and testing data. Note that you should specify the "run_dir" in "configs/exp_interpolation.yaml"
python hydra_run.py --multirun \
experiments=exp_interpolation \
datasets@dataset_spec=tyrol,pyrenees,tfasr \
model_spec.interpolation=identity \
dataset_spec.test_dataset.scale_min=1
Using the checkpoint of TFASR30 to generate super-resolution DEMs. Note that you should specify the "run_dir" in "configs/exp_test.yaml".
python hydra_run.py --multirun \
experiments=exp_test \
datasets@dataset_spec=tfasr \
dataset_spec.test_dataset.scale_min=2,4 \
exp_name="tfasr_edsrB-ebcf-nearest-pe16_best" \
test_spec.model_pth="${abs_pth}/checkpoints/TFASR30_ebcf-nearest-pe16_best-epoch.pth" \
device='cuda:0'
Similar to Pyrenees:
python hydra_run.py --multirun \
experiments=exp_test \
datasets@dataset_spec=pyrenees \
dataset_spec.test_dataset.scale_min=2,4,6,8 \
exp_name="pyrenees_edsrB-ebcf-nearest-pe16_best" \
test_spec.model_pth="${abs_pth}/checkpoints/Pyrenees_ebcf-nearest-pe16_best-epoch.pth" \
device='cuda:0'
and TFASR30to10:
python hydra_run.py --multirun \
experiments=exp_test \
datasets@dataset_spec=tfasr_30to10 \
dataset_spec.test_dataset.scale_min=3 \
exp_name="tfasr30to10_edsrB-ebcf-nearest-pe16_best" \
test_spec.model_pth="${abs_pth}/checkpoints/TFASR30to10_ebcf-nearest-pe16_best-epoch.pth" \
device='cuda:0'
For calculating metrics, you shoud specify the "gt_dir" and the "sr_dir" in "sr-tif.yaml". Note that the "gt_dir" means the "run_dir" in "configs/exp_interpolation.yaml" but shoud be more specific for the dataset. The "sr_dir" means the generated results of the super-resolution model. Also, it can use a flexible way to define vars in "sr-tif.yaml". More details please refer to the useage of "OmegaConf". Now, just run:
python cal_dem_metrics.py
Without the bias prediction:
python hydra_run.py \
experiments=exp_ebcf \
device='cuda:0' \
datasets@dataset_spec=tfasr \
model_spec.interp_mode='none' \
dataset_spec.train_dataset.dataset.repeat=4
Without the pos encoding:
python hydra_run.py \
experiments=exp_ebcf \
device='cuda:0' \
datasets@dataset_spec=tfasr \
model_spec.interp_mode='nearest' \
dataset_spec.train_dataset.dataset.repeat=4
With the pos encoding:
python hydra_run.py \
experiments=exp_ebcf-pe \
device='cuda:0' \
datasets@dataset_spec=tfasr \
model_spec.interp_mode='nearest' \
dataset_spec.train_dataset.dataset.repeat=4 \
model_spec.posEmbeder.spec.n_harmonic_functions=16
python hydra_run.py \
experiments=exp_ebcf-pe \
device='cuda:1' \
datasets@dataset_spec=tfasr_30to10 \
model_spec.interp_mode='nearest' \
model_spec.posEmbeder.spec.n_harmonic_functions=16
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent