Automatic Detection of Marine Litter: A General Framework to Leverage Synthetic Data

The spatial and temporal coverage of spaceborne optical imaging systems are well suited for automated marine litter monitoring. However, developing machine learning-based detection identification algorithms requires large amounts data. Indeed, when it comes to debris, ground validated data is scarce. In this study, we propose a general methodology that leverages synthetic in order avoid overfitting generalizes well. idea utilize realistic models image acquisition generate train the learning algorithms. These can then be used detect pollution automatically on real satellite images. main contribution our study showing trained simulated successfully transferred real-life situations. We present components framework, modeling satellites debris proof concept implementation macro-plastic with Sentinel-2 case generated dataset (more than 16,000 pixels debris) composed seawater, plastic, wood Random Forest classifier it. This classifier, tested images, discriminates from thus proving effectiveness approach paving way even more representative simulation models.