FloatingObjects Dataset - elena-andreini/TriesteItalyChapter_PlasticDebrisDetection GitHub Wiki

FloatingObjects Dataset Card

Dataset Name:

FloatingObjects Dataset

Description:

The FloatingObjects dataset is a specialized collection of Sentinel-2 satellite images focused on detecting and categorizing floating objects in marine environments, including plastic debris, seaweed, and organic matter. The dataset supports machine learning applications aimed at distinguishing floating plastics from other materials and improving the accuracy of marine pollution monitoring.

Paper: https://doi.org/10.5194/isprs-annals-V-3-2021-285-2021

Key Characteristics:

Satellite Mission: Sentinel-2
Data Type: Multispectral imagery from Sentinel-2 satellites

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

Atmospheric Correction:

  • Applied: No
  • Algorithm: NA

Bands : L1C data has 13 bands while L2A data has 12 bands

Spatial Resolution: 60m (B1, B9, B10), 10m (B2, B3, B4, B8), 20m (B5, B6, B7, B8A, B11, B12)

Temporal Coverage:

  1. Data availability: Since October 2016 globally
  2. Revisit time: 5 days for the entire Sentinel-2 constellation (Sentinel-2A and Sentinel-2B combined)
  3. Temporal resolution: 10 days at the equator with one satellite, 5 days with both satellites under cloud-free conditions and 2-3 days at mid-latitudes

Geographic Coverage: Covers Geographically diverse regions from North America, South America, East Africa, Europe, Egypt, also including Mediterranean Sea.

Labels/Annotations:

  • Type: LineString geometries in shapefiles
  • Classes: Floating Object (1 class)

Debris-Free Scenes: No

Relevance to Project:

  1. Include a scene with the Venice area (Italy)
  2. High-resolution Sentinel-2 spectral data for machine learning model training
  3. Multi-seasonal coverage allows for analysis of seasonal plastic accumulation

Access: https://github.com/ESA-PhiLab/floatingobjects

Additional Notes:

  1. Includes various marine environments (coastal, offshore, open sea).
  2. The dataset is suitable for training deep learning models (CNNs, U-Nets, transformers) for segmentation.
  3. The labels in the form of coarse hand-annotated lines contain some inherent noise

Next Steps:

  1. Download the Data Set.
  2. Split the included Sentinel-2 imagery in small patches
  3. Raster shapefile to produce binary masks