LitterLines is a ready-to-analyze dataset for detecting marine litter windrows (MLWs) in PlanetScope imagery, consisting of 29 annotated scenes with over 2.4 million labeled pixels. Designed for machine learning applications, it enables the development and fine-tuning of models for MLW detection, supporting improved monitoring of marine litter from satellite-based sources.

• The dataset consists of 29 PlanetScope scenes, acquired from Dove-C (22 scenes), SuperDove (6 scenes), and Dove-R (2 scenes) satellites.
• Imagery is provided as Analytic Ortho Scene products with at-sensor radiance.
• Conversion to top-of-atmosphere (TOA) reflectance can be performed using the provided repository code.
• Data was collected between September 2024 and February 2025.
• The dataset contains 1,016 image patches (256×256 px) derived from 966 line annotations, using high-resolution PlanetScope imagery.

The dataset follows a hierarchical structure:

LitterLines/
│── 20240115_Algeria/
│   ├── 100e/
│   │   ├── 20180123_100009_100e.json
│   │   ├── 20180123_100009_100e_3B_AnalyticMS_clip.tif
│   │   ├── 20180123_100009_100e_3B_AnalyticMS_metadata_clip.xml
│   │   ├── 20180123_100009_100e_3B_udm2_clip.tif
│   │   ├── 20180123_100009_100e_metadata.json
│   ├── 1018/
│   │   ├── 20180123_095909_1018.json
│   │   ├── 20180123_095909_1018_3B_AnalyticMS_clip.tif
│   │   ├── 20180123_095909_1018_3B_AnalyticMS_metadata_clip.xml
│   │   ├── 20180123_095909_1018_3B_udm2_clip.tif
│   │   ├── 20180123_095909_1018_metadata.json
│
│── 20201202_Egypt/
│   ├── 2264/
│   │   ├── 20201202_075858_57_2264.json
│   │   ├── 20201202_075858_57_2264_3B_AnalyticMS_clip.tif
│   │   ├── 20201202_075858_57_2264_3B_AnalyticMS_metadata_clip.xml
│   │   ├── 20201202_075858_57_2264_3B_udm2_clip.tif
│   │   ├── 20201202_075858_57_2264_metadata.json

• metadata.json: contain acquisition details for each scene.
• AnalyticMS.tif: contain 4-band GeoTIFF's (RGB and NIR) bands.
• AnalyticMS_metadata.xml: contain conversation factors for TOA reflectance
• udm2.tif: Usable Data Mask indicating valid pixels.
• Additional metadata for all samples is compiled in doc/litterlines_metadata.xlsx.

• For detailed documentation on data retrieval and annotation, please contact Joost van Dalen.
• Preprocessing and model training can be facilitated using the ready-to-analyze code in the repository.

# Install LitterLines and its dependencies
pip install git+ssh://git@github.com/geoJoost/LitterLines.git

# Create and activate conda environment
conda create --n litterlines
conda activate litterlines

# Install dependencies
# Replace with appropriate PyTorch version

conda env update --f environment.yml

The LitterLines dataset is designed for marine litter windrow detection, containing 1,016 image patches (256×256 px). The datset is structured through a dataloader for ready-to-use machine learning applications, making it suitable for:

• Fine-tuning pre-trained deep neural networks
• Training machine learning models models
• Analyzing spectral signatures of marine litter windrows

Below is a selection of image patches, visualized using RGB, NDVI, and RAI (Rotation-Absorption Index). The last column shows the annotation label used for training:

The retrieval pipeline is modular, allowing for easy expansion and adaptation to different regions and sensor types.