🏔️ Diff-DEM: A Diffusion Probabilistic Approach to Digital Elevation Model Void Filling

🎯 Purpose

Diff-DEM is a deep learning-based tool designed to fill voids in Digital Elevation Models (DEMs) using a diffusion probabilistic model.
It divides large georeferenced DEMs and their corresponding mask images into smaller tiles, processes them using a pretrained model, and stitches the results into full georeferenced output DEM.

---

📥 Input Format

To run inference, you need:

• ✅ An input DEM image (--input_img)
• 🔄 A binary mask image (--input_mask) - OPTIONAL (will be auto-generated if not provided)
• ✅ A pretrained model checkpoint (--resume_state)
• ✅ Number of diffusion inference steps (--n_timestep, e.g. 512)
• ✅ (Optional) Tile overlap in pixels (--tile_overlap) to reduce tiling artifacts

Example Commands:

# With provided mask (traditional approach)
python run.py -p test -c config/dem_completion.json \
  --resume_state ./pretrained/760 \
  --n_timestep 512 \
  --input_img "Path_to_Input_DEM.tif" \
  --input_mask "Path_to_Mask.tif" \
  --tile_overlap 0

# Without mask (auto-generation from nodata values)
python run.py -p test -c config/dem_completion.json \
  --resume_state ./pretrained/760 \
  --n_timestep 512 \
  --input_img "Path_to_Input_DEM.tif" \
  --tile_overlap 0

---

🧠 Internal Workflow

1. Temporary Directory Creation A temp folder is created to hold intermediate tiles and results.

2. 🆕 Preprocessing: Automatic Mask Generation (NEW)

   • If no --input_mask is provided, automatically generates nodata mask from input TIF
   • Detects nodata values (defined nodata, NaN, infinity, extreme outliers)
   • Creates binary mask (255=nodata, 0=valid data) for inpainting workflow
   • Saves auto-generated mask to temporary directory

3. Tiling

   • Input DEM and mask are split into 128×128 tiles.
   • If --tile_overlap is provided (e.g., 64), tiles overlap with neighbors by that many pixels.

4. FList Creation Tile paths are written into .flist files and fed to the model.

5. Model Inference

   • Model input remains 128×128, regardless of overlap.
   • With overlap, each pixel may be predicted multiple times (e.g., 50% overlap → predicted 3+ times).
   • Tiles values are converted into -1 to 1 ranges for model input/inference.
   • Tile's min_max values are saved using which the predicted images pixel values are scaled back to original 32bit dem pixel values.

6. Output Stitching

   • Predicted tiles are merged back into a full DEM.
   • Overlapping predictions are averaged, producing smooth transitions.
   • Original georeferencing and metadata are preserved.

---

🔍 Tile Overlap: Why It Matters

Abrupt changes and seams may appear between tiles if no overlap is used. This happens because:

• The model only “sees” one tile at a time.
• Masks often span across tile boundaries.
• No context beyond the current tile leads to discontinuities.

Option	Description
No --tile_overlap	Fast but may cause visible seams
With --tile_overlap	Slightly slower, but smoother and seamless outputs

📌 Note: Overlap introduces some redundancy but improves quality.

---

🖥️ Environment Setup

conda env create -f environment.yml
conda activate Diff-DEM

---

📦 Dataset & Pretrained Model

• 🔽 Download: Google Drive Link
• 📁 Place dataset under: Diff-DEM/dataset/norway_dem/ or give input images
• 📁 Place pretrained models at: Diff-DEM/pretrained/

---

⚙️ Command Line Arguments

python run.py -p test -c config/dem_completion.json \
  --resume_state ./pretrained/760 \
  --n_timestep 512 \
  --input_img "Input Image Path" \
  --input_mask "Input Mask Path" \
  --tile_overlap 20

Argument	Description
-p, --phase	train or test (default: test)
-c, --config	JSON config file
-b, --batch	Batch size (per GPU, under development)
--resume_state	Path to model checkpoint
--input_img	Input DEM (GeoTIFF)
--input_mask	🆕 OPTIONAL Binary mask image (auto-generated if not provided)
--tile_overlap	Overlap in pixels for tiling (default: 0)
--nodata_value	Value treated as NoData in input image for mask creation
--scale_factor	Rescale output (default: 1)
--keep_temp	Keep temporary tile outputs
--gpu_ids	GPU IDs to use (e.g., 0, 0,1)
--output_dir_name	Custom output directory name
--preprocess_type	Optional pre-processing flag
--data_root	Path to .flist of DEMs (batch mode)
--mask_root	Path to .flist of masks
--out_type	Output datatype (default: float32)
--use_color_map	Enable color map for TensorBoard
--sample_num	Number of samples to generate
--debug	Enable debug mode
--port	Visualization port (default: 21012)

---

📈 Inference Modes

Standard .flist Mode (Batch)

python run.py -p test -c config/dem_completion.json \
  --resume_state ./pretrained/760 \
  --n_timestep 512 \
  --data_root ./dataset/norway_dem/benchmark/benchmark_gt.flist \
  --mask_root ./dataset/norway_dem/benchmark/mask_64-96.flist

Large Single Image Mode (With Tiling)

# With provided mask
python run.py -p test -c config/dem_completion.json \
  --resume_state ./pretrained/760 \
  --n_timestep 512 \
  --input_img "Input_DEM.tif" \
  --input_mask "Input_Mask.tif" \
  --tile_overlap 64

# 🆕 Without mask (auto-generated)
python run.py -p test -c config/dem_completion.json \
  --resume_state ./pretrained/760 \
  --n_timestep 512 \
  --input_img "Input_DEM.tif" \
  --tile_overlap 64

---

🧪 Training

python run.py -p train -c config/dem_completion.json

Check training progress:

tensorboard --logdir experiments/train_dem_completion_XXXXXX_XXXXXX

---

🧪 Metric Evaluation

python data/util/tif_metric.py \
  --gt_tif_dir ./dataset/norway_dem/benchmark/gt \
  --mask_dir ./dataset/norway_dem/benchmark/mask/128-160 \
  --algo_dir ./experiments/Diff-DEM/128-160/results/test/0 \
  --normalize

---

📝 Known Issues

• Tile artifacts: Without --tile_overlap, visible seams may occur due to lack of tile context.
• Batch processing (-b): Batch option was implemented in source code but throws error therefore testing/development is underway.
• Model Input Size: Always expects 128×128 inputs. Overlap is handled internally—model is agnostic to it.

---

🙏 Acknowledgements

• Palette: Image-to-Image Diffusion Models
• Gavriil's DEM Dataset
• Norwegian Mapping Authority

---

📚 Citation

@article{lo2024diff,
  title={Diff-DEM: A Diffusion Probabilistic Approach to Digital Elevation Model Void Filling},
  author={Lo, Kyle Shih-Huang and Peters, Jörg},
  journal={IEEE Geoscience and Remote Sensing Letters},
  year={2024},
  publisher={IEEE}
}