MethaneSegNet: Methane Plume Segmentation via Multi-scale SwinTransformer-Based Network

Methane emissions are a major contributor to near-term climate warming, requiring accurate detection and monitoring. Satellite remote sensing offers scalability but faces key challenges, including small-pixel plume localization, spectral ambiguity. We propose MethaneSegNet, an attention-based encoder-decoder built on a Swin-Transformer backbone, enhanced by a Multi-scale Aggregation module and Spatial-Transformer Convolutions (STN-Conv). MethaneSegNet tackles these challenges through: (1) small-pixel detection via STN-Conv for fine-grained localization; (2) adaptive multi-modal fusion using spatial and channel attention. On public benchmarks, STARCOP and \(\text {CH}_4\) Net/Sentinel-2, MethaneSegNet outperforms prior methods with gains of up to 4.43% and 5.47% in Precision, demonstrating strong potential for methane plume segmentation and environmental monitoring.