In order to meet the 1.5 °C temperature control target of the Paris Agreement and the implementation requirements of carbon tariff policy, this paper builds a multi-modal remote sensing monitoring framework for carbon neutrality, breaking through the technical bottlenecks of traditional methods in emission traceability accuracy (error > 5%) and dynamic timeliness of carbon sinks (quarterly update). The framework design includes: a multi-modal spatio-temporal fusion model integrating Landsat8, Sentinel-5P TROPOMI and ERA5 data, and using the radiative transfer equation to correct the scale difference, the spatial positioning accuracy of point source emissions is less than 1 km; A dynamic carbon sink assessment algorithm based on the coupling model of NDVI and soil organic carbon was proposed. The parameter sensitivity (NDVI contribution 62.3%) was optimized by machine learning to quantify the interannual change rate of carbon sink. Edge computing monitoring system based on federated learning architecture, realizing hourly data update and response delay < 2 h. The multi-dimensional test verification of the Beijing-Tianjin-Hebei Demonstration zone showed that the F1-score of the industrial point source reached 0.89 (23.6% higher than that of the traditional method), the RMSE of the carbon sink decreased to 0.34 MtC/km2 (65% less error), and the system covered 12 provinces and administrative regions and identified 1235 hidden emission sources. Significant contribution to research innovation: pioneered the three-level fusion paradigm of “radiative transfer, deep learning and federal computing” to support accurate carbon tariff accounting; Develop a dynamic monitoring system for subkilometer carbon sinks to meet the needs of global carbon budget verification under the Paris Agreement; Prototype a scalable global monitoring network to provide the technical tools to achieve UNFCCC goals.