Environmental semantic clustering-guided multimodal fusion for enhanced interpretability in methane concentration prediction

over a 20-year timeframe. Accurately predicting the spatiotemporal distribution of methane concentrations, particularly near industrial sources, is essential for effective environmental monitoring and provides a critical foundation for subsequent emission source identification. This study introduces a novel Spatial-Temporal Cross-Attention Network (ST-CAN) to address the challenge of fusing sparse, high-frequency ground-based observations with spatially extensive but temporally infrequent satellite imagery. Using the Athabasca oil sands region as a case study, ST-CAN incorporates three synergistic innovations that work together to address data fusion challenges: (1) wavelet decomposition transforming high-frequency ground measurements into multi-scale temporal features capturing both long-term trends and short-term emission events; (2) environmental semantic clustering that identifies distinct atmospheric patterns from these wavelet features, providing interpretable contextual labels; and (3) a bidirectional cross-attention mechanism where these semantic cluster labels dynamically guide how ground temporal features query and fuse with satellite spatial information, adaptively prioritizing relevant features based on identified environmental states. The model is designed to leverage time-dense ground data to enhance the temporal resolution of weekly satellite-derived concentration maps, generating high-fidelity spatially representative methane concentration predictions by integrating information from four spatially distributed monitoring stations and satellite imagery, capturing regional-scale atmospheric dynamics. Extensive evaluations demonstrate ST-CAN significantly outperforms all the baseline models in predictive accuracy and robustness. The bidirectional mechanism notably improves interpolation during satellite data gaps, mitigating cloud cover and data sparsity challenges. By combining interpretability with advanced AI techniques, ST-CAN provides a transparent and scalable framework for high-resolution methane concentration modelling, advancing environmental monitoring capabilities and supporting targeted climate mitigation efforts.