Satellite-based marine pollution monitoring for African coastal energy infrastructure: Transformer semantic segmentation of Sentinel-2 multispectral imagery on the MADOS benchmark

Abstract Chronic oil pollution and marine litter around coastal energy infrastructure remain poorly monitored in many African waters, where field access is limited and cloud cover restricts routine observation. Publicly available Sentinel-2 imagery offers a practical basis for operational monitoring, yet most published detectors rely on purpose-built architectures whose advantages over general-purpose models are unclear. In this study, two widely adopted transformer segmentation architectures, SegFormer-B2 and Mask2Former, were evaluated on the Marine Debris and Oil Spill (MADOS) dataset of 174 Sentinel-2 scenes with roughly 1.5 million annotated pixels across 15 classes. Both models were adapted for four-channel red, green, blue and near-infrared input by weight inflation of pretrained filters, trained on the official MADOS splits, and compared against the published MariNeXt baseline using mean intersection over union (mIoU) and per-class F1. SegFormer-B2 reached a mIoU of 0.726 and a macro-F1 of 0.807, and Mask2Former reached 0.623 and 0.714, against 0.643 and 0.760 for MariNeXt. Both transformers detected oil spills (F1 of 0.94 and 0.95 against 0.78) and marine debris (0.81 and 0.75 against 0.51) more accurately than the baseline, but lost ground on sparse floating algae, foam and jellyfish, a pattern consistent with the red-edge and shortwave infrared bands omitted from the four-channel input. A pretrained general-purpose transformer can therefore exceed a purpose-built multispectral model on aggregate accuracy and on the pollutant classes of greatest operational interest, offering a transferable reference point for satellite-based pollution monitoring around African coastal energy infrastructure.