MTU-Former: A Multitask Unified Transformer Framework for Global Methane Monitoring With Million-Scale Multispectral Imagery

Abstract

Accurate monitoring of methane emissions is critical for addressing global climate change. While Sentinel-2 short-wave infrared (SWIR) imagery offers high spatiotemporal resolution for monitoring the dynamics of methane emissions, its complex background noise challenges automated data analysis. To address this, we propose a multitask unified transformer (MTU-Former) framework designed to simultaneously detect, segment, and quantify methane plumes by leveraging multitask synergy to enhance quantification performance. Central to our approach is the novel development of a mask-guided cross-task feature fusion (MG-CTFF) module. By fusing the plume probability map and the retrieved methane column enhancement map through a contribution map and dual physics-aware attention biases, the module guides regression attention toward core plume regions while suppressing background interference. To further resolve task discrepancies, the framework integrates a fast adaptive multitask optimization (FAMO) strategy to balance gradient updates. To support this data-driven approach, we present a large-scale dataset comprising 3 million images, constructed by combining Sentinel-2 data from 12 global regions with weather research and forecasting (WRF)-large eddy simulation (LES)-simulated plumes. Comprehensive experimental results show that MTU-Former achieves a coefficient of determination ( $R^{2}$ ) of 0.9771, a root-mean-square error (RMSE) of 1532.93 kg h−1, and a mean absolute percentage error (MAPE) of 9.55%. In pixel-level recognition, our model achieves a detection accuracy of 99.18%, a Dice coefficient of 0.9151, and an intersection over union (IoU) of 0.8673. Compared with three deep learning models, including MethaNet, Swin-UNet, and Efficient-UNet, the proposed framework reduces the RMSE by 1633.35, 562.75, and 370.47 kg h−1, respectively. Correspondingly, it lowers the MAPE by 10.48%, 4.45%, and 3.09%, while improving $R^{2}$ by margins of 0.0749, 0.0199, and 0.0124. In comparison with the background-corrected integrated mass enhancement (IME) method, MTU-Former reduces the RMSE from 4323.78 to 2019.79 kg h−1 and MAPE from 16.76% to 9.16%, while increasing $R^{2}$ from 0.6411 to 0.9217. Overall, MTU-Former serves as a robust, versatile technical approach for significantly advancing multitask monitoring of global-scale methane emissions.