Self-supervised pre-training enables marine debris detection across sensors

Plastic pollution is a global crisis that needs to be addressed at scale. Current efforts have focused on annotating Sentinel-2 acquisitions for floating debris, and machine learning models are trained to automatically detect such marine debris. However, Sentinel-2 has a revisit time of approximately one week, and images are acquired at 10m resolution, leading to insufficient spatial and temporal resolution for collection efforts. To pave the way for daily monitoring of marine pollution at high spatial and temporal resolution, we propose a cross-sensor approach that allows to transfer the model’s knowledge from Sentinel-2 to daily-acquired PlanetScope images at 3m resolution. Being self-supervised, the proposed method does not require labels for PlanetScope data, therefore relieving the burden of creating a large-scale annotated set of PlanetScope images. Our approach shows its effectiveness over the Marine Debris Detector on a set of images that we annotated for validation purposes. We share the annotations and the model for benchmarking. • Self-supervised framework enabling to detect marine debris without additional labels. • We release a benchmark dataset of PlanetScope images with dense binary labels. • CSM-Debris paves to way for daily marine debris monitoring across scales and sensors.