Super-resolution GAN-enhanced detection of sub-pixel marine plastic debris: Leveraging Sentinel-2 time-series and...

AI-related satellite observation at the superpixel level has remained one of the major challenges for marine plastics because multispectral sensors are mostly lower in spatial resolution. The present research has therefore proposed a technique using the physics-informed deep learning super-resolution-generative adversarial network (SR-GAN) to enhance subpixel marine plastic detection from multi-temporal Sentinel-2 imagery. It utilizes multi-Sentinel-2 MSI spectral bands and plastic-sensitive indices to improve spatial details and spectral separability of floating plastic features in complex coastal waters. The primary objective of this paper is to improve spatial representation of sub-pixel marine plastics using super-resolution deep learning, detection reliability, however, to be enhanced using physics-based spectral constraints inherent to the GAN architecture. Area study included the Balikpapan Coastal Region using multi-temporal Sentinel-2 time series for checking assessment of surface conditions over time at sea, whereby model performance was assessed from several coastal sub-regions. Detection by SRGAN was at a high confidence level of 96.8% for Balikpapan, low probability of false alarm (0.04), and low probability of missing a plastic object (0.03). The super-resolution enhancement factor was found to be 2.8 times, with 97.4% testing confidence in Muarasempum, while 99.1% testing confidence in Sesulu had no missed detections, showing strong generalization and minimal over-detection. There were super-resolution enhancements across all areas, ranging from 2.1× to 2.8×, indicating robustness of the above-mentioned approach. It is evident that physics-informed SR GANs would enhance subpixel marine plastics mapping tremendously and would scale towards operational solutions for the monitoring of coastal plastics.