The invention discloses a multi-source-driven CO2 concentration high-temporal-spatial-resolution prediction method based on a dynamic
diffusion model, and the method integrates the
satellite observation of OCO-2, GOSAT and the like, ERA5 meteorological data,
vegetation index NDVI,
population activity and other characteristics on the basis of DCRNN, and achieves high-precision prediction through physical prior and a dynamic graph structure. The method comprises the following steps: unifying multi-
source data to a
regular grid; an adjacent matrix is dynamically generated based on the geographic position and the
wind field to serve as the input of the improved DCRNN, and the
diffusion relation of COs in the
wind direction is described; multi-source node features are input into the improved DCRNN, and physical constraints such as
mass conservation and
convection consistency are introduced to improve the reliability of a result; a graph-to-grid decoding and super-resolution technology is adopted to reconstruct prediction to a higher resolution level by level, and pixel-level uncertainty can be output. According to the method, multi-source observation and meteorological driving information are fully utilized, the
satellite observation blank is effectively supplemented, the
fineness and continuity of a COconcentration product are improved, and the method is suitable for carbon emission accounting,
carbon source and sink evaluation and environmental decision support.