Sentinel-2 Forest Browning Monitoring

Code for Country-wide, high-resolution monitoring of forest browning with Sentinel-2 (ISPRS Congress 2026)

👥 Authors: Samantha Biegel, David Brüggemann, Francesco Grossi, Michele Volpi, Konrad Schindler, Benjamin Stocker
🌐 Website: samanthabiegel.github.io/forestbrowning
📄 Paper: arXiv 2604.02074
🗺️ Demo: forest-monitoring.org

This project generates Switzerland-wide, 10 m resolution NDVI anomaly maps from Sentinel-2 imagery to monitor forest browning events (drought stress, beetle outbreaks, storm damage, fire, and clear-cuts). A neural network learns the expected seasonal vegetation cycle per pixel; deviations from this expectation are flagged as anomalies.

---

Repository contents

s2-forest-browning-monitoring/
├── src/forest_browning/
│   ├── config.py              # path constants (configured via env var)
│   ├── dataset.py             # dataset utilities
│   ├── mlp.py                 # autoencoder architecture
│   ├── train.py               # model training
│   ├── inference.py           # anomaly score generation
│   ├── shuffle_train_data.py  # pre-shuffle training data
│   ├── rechunk_output.py      # reformat output for spatial access
│   └── data_processing/       # 12-step dataset creation pipeline
├── notebooks/
│   └── plot_results.ipynb     # reproduces all figures → figs/
├── data/
│   └── event_polygons/        # labelled disturbance event polygons
├── checkpoints/
│   └── encoder.pt             # pre-trained model checkpoint
├── tests/
├── pyproject.toml
├── uv.lock
└── README.md

---

Requirements

• Python 3.12 or later
• uv
• GDAL CLI tools (gdalwarp, gdaldem, gdal_calc.py)
• TauDEM (d8flowdir, aread8) with MPI

---

Installation

GDAL

Ubuntu/Debian:

sudo apt install gdal-bin

macOS:

brew install gdal

Verify:

gdalwarp --version

TauDEM

Step 6 of the data pipeline uses TauDEM for hydrological feature computation. TauDEM is a compiled binary and requires an MPI implementation.

Ubuntu/Debian:

sudo apt install libgdal-dev libopenmpi-dev cmake git
git clone https://github.com/dtarb/TauDEM.git
cmake -S TauDEM/src -B TauDEM/build && cmake --build TauDEM/build -j4
sudo cmake --install TauDEM/build

macOS:

brew install open-mpi cmake
git clone https://github.com/dtarb/TauDEM.git
cmake -S TauDEM/src -B TauDEM/build && cmake --build TauDEM/build -j4
sudo cmake --install TauDEM/build

Verify:

which d8flowdir && mpiexec --version

Python environment

git clone git@github.com:SamanthaBiegel/s2-forest-browning-monitoring.git
cd s2-forest-browning-monitoring
uv sync --group dev
source .venv/bin/activate

Install the pre-commit hooks:

pre-commit install

---

Configuration

Before running any pipeline step, set the environment variable FOREST_BROWNING_DATA_DIR to point to your local data storage:

export FOREST_BROWNING_DATA_DIR=/your/local/data/dir

Note: two external datasets must be downloaded manually and placed in FOREST_BROWNING_DATA_DIR:

• Tree species map: request access via https://www.envidat.ch/#/metadata/tree-species-map-of-switzerland and place the obtained .tif raster in FOREST_BROWNING_DATA_DIR.
• Habitat map: download from https://www.envidat.ch/#/metadata/the-habitat-map-of-switzerland-v1-1 and place habitatmap_v1_1_20241025.tif in FOREST_BROWNING_DATA_DIR.

---

Dataset creation

The dataset building pipeline consists of 12 steps orchestrated by a runner module. The individual step scripts are in src/forest_browning/data_processing/ with digit prefixes (e.g., 1_extract_swisstopo_dataset.py); the runner invokes them sequentially and exits if any step fails.

Activate the virtual environment first:

source .venv/bin/activate

Option 1: Use the Python module

python -m forest_browning.data_processing.pipeline

Option 2: Run a single step manually

To run one step in isolation:

python src/forest_browning/data_processing/1_extract_swisstopo_dataset.py

Output dataset

The pipeline produces two Zarr datasets:

File	Chunking	Use case
ndvi_dataset_temporal.zarr	(num_forest_pixels, num_timesteps)	Training and inference
ndvi_dataset_spatial.zarr	(num_timesteps, num_forest_pixels)	Fast per-day map retrieval

A forest mask (forest_mask.npy) maps between the 1-D pixel index used in the datasets and the original 2-D spatial grid.

---

Training and inference

Activate the virtual environment first:

source .venv/bin/activate

Step 1 – Pre-shuffle the training dataset:

python -m forest_browning.shuffle_train_data \
    --input_zarr /path/to/ndvi_dataset_temporal.zarr \
    --output_zarr /path/to/ndvi_dataset_filtered_shuffled.zarr

Step 2 – Train the neural network:

python -m forest_browning.train \
    --data_path /path/to/ndvi_dataset_filtered_shuffled.zarr \
    --output_dir /path/to/work/dir

Step 3 – Run inference to generate anomaly scores:

This step will add the anomaly scores as an additional layer to the supplied zarr file. Adapt --encoder_path path if you want to run a model trained in step 2.

python -m forest_browning.inference \
    --encoder_path checkpoints/encoder.pt \
    --data_path /path/to/ndvi_dataset_temporal.zarr \
    --output_path /path/to/ndvi_dataset_temporal.zarr

Step 4 – Rechunk output for spatial access:

python -m forest_browning.rechunk_output \
    --source_zarr /path/to/ndvi_dataset_temporal.zarr \
    --target_zarr /path/to/ndvi_dataset_spatial.zarr

---

Reproducing figures

Open and run notebooks/plot_results.ipynb. Figures are saved to figs/.

---

Citation

If you use this code or data in your research, please cite:

@proceedings{biegel2025forestbrowning,
  title   = {Country-wide, high-resolution monitoring of forest browning with {Sentinel-2}},
  author  = {Biegel, Samantha and Br{\"u}ggemann, David and Grossi, Francesco and Volpi, Michele and Schindler, Konrad and Stocker, Benjamin},
  booktitle = {ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences},
  year    = {2026},
}

---

License

This project is licensed under the terms of the LICENSE file.