MADOS Dataset Card - elena-andreini/TriesteItalyChapter_PlasticDebrisDetection GitHub Wiki

MADOS Dataset Card

Dataset Name:

MADOS (Marine Debris and Oil Spill) Dataset

Description:

The MADOS dataset is a multi-class marine pollution dataset designed for detecting floating marine debris and oil spills using Sentinel-2 imagery. It provides pixel-wise segmentation annotations, distinguishing between different types of floating pollutants in oceanic environments. The dataset is useful for training machine learning and deep learning models for plastic debris detection, while also enabling differentiation between oil spills, organic matter, and clean water.

Key Characteristics:

Satellite Mission: Sentinel-2
Data Type: Sentinel-2 imagery. MADOS is composed of 174 scenes, which are segmented into 2803 tiles. Each tile measures 240x240 pixels

Imagery Level: Level-1C and Level-2A

Atmospheric Correction:

  • Applied: Yes
  • Algorithm: ACOLITE

Bands : All bands except B10

Spatial Resolution: 60m (B1, B9), 10m (B2, B3, B4, B8), 20m (B5, B6, B7, B8A, B11, B12)

Temporal Coverage:

  1. Sentinel-2 acquisitions from 2016 – 2022
  2. Multi-seasonal coverage to study plastic accumulation variations
  3. Includes different wave & weather conditions

Geographic Coverage:

  1. Global dataset, including key locations in the Mediterranean & Italian seas
  2. Focuses on plastic pollution hotspots & oil spill incidents
  3. Includes coastal, offshore, and open-sea environments

Labels/Annotations:

  • Type: Pixel-wise segmentation
  • Classes:
    1. Plastic debris
    2. Oil spills
    3. Algae & organic materials
    4. Clean water (debris-free scenes included)

Debris-Free Scenes: Yes

Relevance to Project:

  1. Differentiates plastic debris from oil spills, reducing false positives in classification.
  2. Multi-seasonal dataset ensures models generalize across different conditions.
  3. High-quality pixel-wise annotations make it suitable for deep learning-based segmentation models.
  4. Combines Sentinel-2 Level-1C & Level-2A data, providing flexibility in preprocessing.

Access: https://zenodo.org/records/10664073

Additional Notes:

  1. Cloud masking may be required (using Sentinel-2 QA60 or SCL bands).
  2. Oil spill & plastic debris differentiation can be challenging, requiring spectral analysis.
  3. Spectral indices such as NDPI, FDI, and IndexMP can improve plastic detection accuracy.
  4. Annotations may need filtering if only plastic debris (and not oil spills) is required.

Next Steps:

  1. Apply cloud masking (Sentinel-2 QA60 band or SCL band).
  2. Data Filtering & Annotation Refinement
  3. Separate debris-free vs. polluted scenes for better training.
  4. Compute spectral indices for plastic detection.