Denoising as a strategy for Detecting Ocean Plastic Pollution with few labels

Marine Litter is a growing ecologic concern that needs to be addressed at a global scale. Even though satellite and meteorological data is abundant, continuously monitoring marine litter aggregated in floating patches close to coastlines is still an open technological question and an ongoing area of research. IIn the context of the ADOPT project between EPFL, the Ocean Cleanup and the Swiss Data Science Center,we propose a deep learning-based detection module to identify aggregations of floating debris from remote sensing data. As remote sensing images are widely available, but only a small portion is annotated, we demonstrate that the unsupervised training of a Denoising AutoEncoder (DAE) on Sentinel-2 marine scenes allows the learning of discriminative features for the semantic segmentation task of floating debris identification. Fine-tuning the denoising autoencoder on only 10% of the labeled dataset leads to a performance on par with that of a supervised model trained on the full dataset. The development of a label-efficient model reduces the need for annotated data and suggests that less annotated images of higher quality are better suited than large but potentially noisy datasets.