Automatic Detection of Floating Marine Debris Using Multi-spectral Satellite Imagery,

07/08/2022

Marine plastic pollution represents a maritime environmental emergency that needs to be addressed. Floating plastic debris must be detected, captured, and removed from the ocean, in order to preserve such a fragile ecosystem. In this work, it is shown that floating plastic debris are not only detectable but also distinguishable from other floating materials, such as driftwood, seaweed, sea snot, sea foam, and pumice, in optical data from the European Space Agency (ESA) Sentinel-2 satellites, using a supervised learning method trained with data compiled from published works and complemented by some manual interpretation of satellite images. The proposed model, an Extreme Gradient Boosting (XGBoost) trained with seven spectral indices and two spectral bands, successfully classified 98% of the pixels that contained floating plastic debris in coastal waters. Additionally, due to the need for more floating plastic data in the training dataset, synthetic data were generated through a Wasserstein Generative Adversarial Network (WGAN). A supervised model trained only with synthetic data successfully classified plastic pixels with an accuracy of 91%. Finally, to build a system that provides reliable results when applied in real-world conditions, an ensemble model that quantifies uncertainty was created. This novel approach correctly classified 79% of the plastic pixels. However, the number of misclassifications decreased significantly compared to the model with the highest accuracy, making it the best option to monitor the ocean.

Authors in the community:

• 

  M

  Miguel Mendes Duarte

  ist425449

Supervisors of this institution:

• 

  L

  Leonardo Azevedo Guerra Raposo Pereira

  ist168809

electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering

eng - English

Instituto Superior Técnico