Accurate monitoring of methane (CH₄) emissions from oil and gas infrastructure is critical for near-term climate change mitigation. While hyperspectral satellites like PRISMA provide high-fidelity data, the optimal deep learning architecture for segmenting faint methane plumes remains undetermined. This study provides the first systematic comparison of convolutional and transformer-based models for this task.
We evaluate four architectures: an attention-enhanced U-Net++, Segformer, and two novel hybrids – SwinV2-UPP (SwinV2 encoder + U-Net++ decoder) and SwinV2-Former (SwinV2 encoder + MLP decoder). Models are trained on 73 manually delineated methane plumes from 34 PRISMA scenes across global oil and gas basins, augmented to 1,446 patches.
SwinV2-UPP achieved the highest F1-score of 0.834, a 4.0% improvement over the convolutional baseline (0.802). Notably, SwinV2-Former achieved state-of-the-art precision of 0.910 – outperforming SwinV2-UPP by 2.5% – demonstrating exceptional false-positive minimization. In zero-shot generalization, SwinV2-UPP maintained the highest F1 (0.725), while SwinV2-Former achieved the best precision (0.857).
We reveal a critical architectural trade-off: SwinV2-UPP balances precision and recall (F1=0.834) for comprehensive plume mapping, whereas SwinV2-Former prioritizes high-confidence detection (precision=0.910) for regulatory applications where false positives are costly. These findings provide a framework for selecting optimal deep learning architectures based on specific methane monitoring objectives.