Skip to content

README.md

Latest commit

 

History

History
187 lines (158 loc) · 8.38 KB

README.md

File metadata and controls

187 lines (158 loc) · 8.38 KB
Raw

ForTy (v1): A Benchmark Dataset for Global Forest Types Mapping

Paper: Not every tree is a forest: benchmarking forest types from satellite remote sensing

ForTy is a new large-scale, multi-modal, and multi-temporal benchmark dataset designed for advancing global FORest TYpes mapping. It comprises 200,000 time series of image patches, each including Sentinel-2, Sentinel-1, climate, and elevation data. The dataset features per-pixel annotations that distinguish between three key forest types (natural forest, planted forest, tree crops) and five other land use/land cover classes. This benchmark aims to support the development of accurate and reliable models critical for efforts like halting deforestation, biodiversity conservation, and compliance with regulations such as the European Union Deforestation Regulation (EUDR).

ForTy (v1) sample locations
ForTy (v1) sample locations.

Accurate differentiation between various forest types (e.g., natural forests, planted forests, and tree crops) is crucial for effective policy-making, conservation strategies, and sustainable forest management. Most existing land use products and benchmarks often categorize all forest areas into a single class or lack the detailed per-pixel annotations needed for precise segmentation tasks on a global scale. ForTy addresses these limitations by providing:

  • Detailed Forest Classes: Distinguishing between natural forests, planted forests, and tree crops.
  • Per-Pixel Labels: Enabling image segmentation tasks at 10 meter resolution.
  • Multi-modal & Multi-temporal Data: Incorporating multi-spectral/optical, polarimetric synthetic aperture radar (SAR), climate, and elevation data over time to capture complex variations.
  • Global Coverage: Covering most land areas with a bias for diverse forest types, leveraging multiple public data sources.

Dataset characteristics

  • Total Samples: 200,000 globally distributed time series of image patches (about 1 TB).

  • Patch Size: Each plot covers a 1280 x 1280 meter area.

  • Input Modalities:

    • Sentinel-2: Multispectral optical imagery (10 bands, 10m and 20m resolution). Seasonal mosaics for 2020.
    • Sentinel-1: Synthetic Aperture Radar (SAR) data (VV and VH polarizations, ascending/descending orbits, 10m resolution). Same temporal extent as Sentinel-2.
    • Climate Data: Monthly climate and water balance variables (~4km resolution, from TerraClimate).
    • Elevation Data: Elevation, slope, and aspect (30m resolution, from FABDEM).

  • Temporal Information: Each time series captures variations at monthly or seasonal cadence.

  • Annotations: Per-pixel segmentation labels.

  • Land Cover Classes (8 classes):

    1. Natural Forest
    2. Planted Forest
    3. Tree Crops
    4. Other Vegetation (shrubland, grassland, cropland)
    5. Water
    6. Ice
    7. Bare Ground
    8. Built Areas
    9. Unknown label is used for pixels with disagreement between source datasets during integration or if unknown.

  • Data Splitting: The dataset is divided into training, validation, and test sets using geographically distinct 100 km x 100 km blocks to reduce spatial autocorrelation and ensure robust model evaluation. Split sizes:

    • train: 160,191
    • validation: 19,661
    • test: 20,148

  • Features in each example: Dimensions are ([Time steps/cadence], Height, Width, [Number of bands/channels]). The temporal and spectral dimensions are optional, depending on the data type.

    Name Dimensions Dtype Description
    id () int64 Sample ID
    lat () float64 Latitude [deg]
    lon () float64 Longitude [deg]
    s2 (4, 128, 128, 10) float32 Seasonal Sentinel-2
    s2_mask (4, 128, 128, 10) uint8 Seasonal Sentinel-2 mask
    s1_asc (4, 128, 128, 3) float32 Seasonal Sentinel-1 ascending
    s1_asc_mask (4, 128, 128, 3) uint8 Seasonal Sentinel-1 ascending mask
    s1_desc (4, 128, 128, 3) float32 Seasonal Sentinel-1 descending
    s1_desc_mask (4, 128, 128, 3) uint8 Seasonal Sentinel-1 descending mask
    elevation (128, 128, 3) float32 Elevation [m], slope, aspect
    climate (4, 1, 1, 14) float32 Seasonal climate variables
    segmentation_labels (128, 128) uint8 Segmentation labels

  • Visualization of some examples

ForTy (v1) examples from different data sources
Each row represents one sample location. Columns from left to right: (1) Sentinel-2 RGB bands, (2) Sentinel-2 SWIR-NIR-Red bands, (3) multi-temporal Sentinel-1 VV winter-spring-summer composition, (4) elevation, (5) labels. Bottom row: color bar for labels.

Data Access

The ForTy v1 dataset is publicly available on Google Cloud Storage (GCS). Current last version is 1.0.0.

The dataset is provided in the TFRecord format. It is split into train, test, and validation sets. Each split is further divided into 1024 shards to facilitate efficient loading and processing. The file naming convention for the shards is forty_v1-{split}.tfrecord-{shard_num:05d}-of-{num_shards:05d} (e.g., forty_v1-train.tfrecord-00000-of-01024).

The data can be downloaded using Cloud SDK, which provides the gcloud utility, with a command like (be careful that it will start a download of 1 TB):

mkdir /tmp/forty_v1
gsutil storage cp -r gs://forest_typology/forty_v1/1.0.0 /tmp/forty_v1/

Data Usage

The simplest way to use the data is via Tensorflow Datasets (TFDS). These python commands demonstrate one way how to access it:

import tensorflow_datasets as tfds
ds = tfds.load("forty_v1", data_dir="gs://forest_typology", try_gcs=True)
# Get a batch of 4 examples (as numpy arrays) for inspection
batch = next(ds["train"].batch(4).as_numpy_iterator())

For more, check the notebook example.

Citing this work

You can cite this work as

@inproceedings{jiang2025:forty-igarss,
  title={Not every tree is a forest: benchmarking forest types from satellite
  remote sensing},
  author={Yuchang Jiang and Maxim Neumann},
  booktitle={Presented at IEEE IGARSS 2025},
  url={https://arxiv.org/abs/2505.01805},
  pages={1-6},
  year={2025}
}

License and disclaimer

Copyright 2025 DeepMind Technologies Limited

The ForTy (v1) dataset has a Creative Commons Attribution ShareAlike 4.0 International License.

All other non-software materials are licensed under the Creative Commons Attribution 4.0 International License (CC-BY). You may obtain a copy of the CC-BY license at: https://creativecommons.org/licenses/by/4.0/legalcode

Unless required by applicable law or agreed to in writing, all software and materials distributed here under the Apache 2.0 or CC-BY licenses are distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the licenses for the specific language governing permissions and limitations under those licenses.

This is not an official Google product.