AN OPEN-SOURCE DATA PIPELINE FRAMEWORK TO DETECT FLOATING MARINE PLASTIC LITTER USING SENTINEL-2 IMAGERY AND MACHINE LEARNING

Plastic accumulation in the ocean is a major global problem that requires an observational system to better inform on the sources and fate of plastic pollution. Here, we present a framework based on open-source satellite data to monitor floating plastic. The data pipeline detects floating plastic and other features (e.g., floating macroalgae) using Sentinel-2 satellite imagery and machine learning. The end-to-end pipeline includes modules for data acquisition, pre-processing, and pixel-based classification using Random Forest or XGBoost models. The models were trained with spectral signatures from events available in literature and show satisfactory accuracies. The framework is applied to different regions and considerations for improvements are presented. The data pipeline allows to detect large enough features that can be suspicious in terms of aggregation of floating plastic litter and therefore be used to alert and inform stakeholders.