Hybrid Deep Learning Approach for Marine Debris Detection in Satellite Imagery Using UNet with ResNext50 Backbone

Marine debris is persistent solid stuff in the water. Oceans include several varieties of organic marine debris, but massive levels of man-made marine trash threaten their biological equilibrium. Manually scanning the ocean for garbage is time-consuming and inefficient, making it uneconomical. Deep learning, which is more efficient than manual methods, is used to detect marine debris in satellite imagery in our work. Deep learning algorithms have been successful in semantic segmentation, however marine debris detection using satellite imagery has been underexplored. The lack of comprehensive marine debris datasets until recently and the complexity of multispectral satellite photos are to blame. Our segmentation method using the UNet architecture and a ResNext50 backbone exceeds the existing state of the art on the Marine Debris Archive Dataset (MARIDA), a dataset of 11 band sentinel 2 Satellite image patches. The hybrid solution combines ResNext50's increased feature extraction with UNet's global and local context preservation, which is crucial in satellite photos of floating bodies due to marine debris' movement pattern. We achieved benchmark mean pixel accuracy, IoU, and F1 scores. We achieved an 88% recall, a 10% improvement over the state of the art, in categorizing marine trash pixels in photos. This work attempts to advance deep learning algorithms for remote sensing and move closer to cleaner oceans.