osmarluiz/coffee-segmentation-dataset

Coffee Plantation Segmentation Dataset

Semantic segmentation dataset for coffee plantation and eucalyptus mapping in the Brazilian Cerrado using Sentinel-1 SAR time series.

Associated paper: Exploiting Convolutional and Transformer Networks for Coffee Plantation Mapping in Brazil Using Polarimetric-Spectral-Temporal Data from Sentinel-1 and Planet NICFI Time Series (under review)

Code: github.com/osmarluiz/coffee-segmentation

---

Dataset Description

This dataset contains 2,400 densely annotated tiles (512 × 512 pixels) from Patrocinio, Minas Gerais — Brazil's largest coffee-producing municipality, located in the Cerrado Mineiro region.

Contents

File	Description	Size
vh_train.tar.gz	Sentinel-1 VH training images (2000 tiles)	25 GB
vh_val.tar.gz	Sentinel-1 VH validation images (200 tiles)	2.5 GB
vh_test.tar.gz	Sentinel-1 VH test images (200 tiles)	2.5 GB
vv_train.tar.gz	Sentinel-1 VV training images (2000 tiles)	25 GB
vv_val.tar.gz	Sentinel-1 VV validation images (200 tiles)	2.5 GB
vv_test.tar.gz	Sentinel-1 VV test images (200 tiles)	2.5 GB
masks.tar.gz	Ground truth masks (all splits)	7 MB

Modalities

Modality	Bands	Format	Description
VH	30	GeoTIFF (.tiff)	Sentinel-1 VH polarization, 30 temporal acquisitions (2021)
VV	30	GeoTIFF (.tiff)	Sentinel-1 VV polarization, 30 temporal acquisitions (2021)
Masks	1	PNG (.png)	Ground truth segmentation masks

Classes

Value	Class	Color
0	Background	Black
2	Coffee	Green
3	Eucalyptus	Yellow

Note: During training, labels are remapped to 0, 1, 2. See the training code for details.

Data Splits

Split	Tiles	Usage
Train	2,000	Model training
Val	200	Validation during training
Test	200	Final evaluation

---

SAR Preprocessing

Sentinel-1 images were acquired in Ground Range Detected (GRD) format, Interferometric Wide Swath mode, and preprocessed using SNAP:

1. Apply orbit file
2. Radiometric calibration (to backscatter)
3. Doppler terrain correction (SRTM)
4. Conversion to decibels (dB)
5. Savitzky-Golay temporal filtering (window=9, polynomial order=2)

---

Planet NICFI Data

The original study also used Planet NICFI optical imagery (48 bands: B, G, R, NIR × 12 months). Due to the NICFI license restrictions, Planet data cannot be redistributed here.

To reproduce the full paper results (Combined dataset, 108 bands):

1. Obtain Planet NICFI mosaics from the NICFI archive
2. Use prepare_combined.py from the code repository to combine all modalities

---

Usage

Download and extract

# After downloading the files:
tar xzf vh_train.tar.gz
tar xzf vh_val.tar.gz
tar xzf vh_test.tar.gz
tar xzf vv_train.tar.gz
tar xzf vv_val.tar.gz
tar xzf vv_test.tar.gz
tar xzf masks.tar.gz

Expected directory structure

DATASET/
├── VH_GT6/
│   ├── image_train/    # 2000 .tiff files
│   ├── image_val/      # 200 .tiff files
│   └── image_test/     # 200 .tiff files
├── VV_GT6/
│   ├── image_train/
│   ├── image_val/
│   └── image_test/
└── Planet_GT6/
    ├── class_train/    # 2000 .png masks
    ├── class_val/      # 200 .png masks
    └── class_test/     # 200 .png masks

Train a model

git clone https://github.com/osmarluiz/coffee-segmentation.git
cd coffee-segmentation
pip install -r requirements.txt
python train.py --dataset VH --model Segformer --encoder efficientnet-b7 --data-dir /path/to/DATASET

---

Citation

@article{carvalho2025coffee,
  title={Exploiting Convolutional and Transformer Networks for Coffee Plantation
         Mapping in Brazil Using Polarimetric-Spectral-Temporal Data from
         Sentinel-1 and Planet NICFI Time Series},
  author={Carvalho, Osmar Luiz Ferreira de and Carvalho Junior, Osmar Abilio de
          and Albuquerque, Anesmar Olino de and Castro Filho, Hugo Cristomo de
          and Moura, Joelma Mendes de and Antony, Dora Silva
          and Silva, Daniel Guerreiro e},
  year={2025},
  note={Under review}
}

License

This dataset is released under CC BY 4.0.