MAM-YOLOv9: A Multiattention Mechanism Network for Methane Emission Facility Detection in High-Resolution Satellite Remote Sensing Images

Abstract

Over 150 countries have signed the Global Methane Pledge, aiming to reduce anthropogenic methane emissions by 30% by 2030. Reducing methane emissions from the energy sector is crucial to achieving this target. The current emission inventories for the energy sector have a spatial resolution of 1 km, suitable for regional-scale methane flux inversion but inadequate for identifying and monitoring point source emissions which is the most important type of anthropogenic methane emissions in the energy sector. To address this issue, we propose a multiattention mechanism, MAM-YOLOv9, for identifying emission facilities in the oil and gas industry, based on YOLOv9. We integrate SimAM and cascaded group attention (CGA) modules into the network, focusing on target objects under complex backgrounds while improving detection accuracy. In addition, we introduce the dynamic convolution module to replace the convolution in the YOLOv9 backbone network, improving computational efficiency and accurate object detection capability. Using submeter-level optical images provided by the high-resolution satellite images, we achieve large-scale monitoring of facility-level emission sources on a regional scale. Experiments demonstrate that our new method achieved SOTA performance, achieving the best results across various metrics compared with the baseline. We also conduct batch detection tasks in Shengli Oilfield, the second-largest oilfield in China, identifying over 38000 emission facilities. Based on the results, we further compile a facility-level methane emission inventory, which can better serve the global efforts for mitigating methane emissions from the oil and gas industry.