Abstract: At present, carbon perception based on greenhouse gas satellite remote sensing technology is gradually becoming a crucial component of new-generation carbon measurement methods. However, accurately extracting carbon emissions data generated by human activities from carbon satellite data represents a key and highly challenging task. In this paper, we propose a novel artificial intelligence algorithm, integrating carbon satellite and electricity emission data, to achieve precise carbon emission measurement. Firstly, we introduce the multimodal data sources used, including carbon satellite and power data, and design corresponding data processing methods. Subsequently, we propose a deep learning method that considers the characteristics of this multimodal data, and construct a data-driven model that reflects the functional relationship among carbon satellite data, power generation data, and carbon source emissions. Finally, based on the carbon concentration remote sensing data from the American OCO-2 carbon satellite and continuous emission monitoring system（CEMS） data from 1 304 American power plants, we validate the effectiveness of the proposed method in the measurement of carbon emissions from power plants.