Abstract: In this study， a high-fidelity synthetic aperture radar （SAR） imaging method is proposed to improve the accuracy of SAR simulation in complex scenarios， using multi-source data fusion and physical modeling. The proposed method is based on the classification of land cover in optical images. First， an improved land cover classification method based on graph theory is adopted to semantically segment the high-resolution RGB optical images. Ground objects are divided into categories such as vegetation， water， soil， and buildings. Meanwhile， each type of ground object is assigned a unique identifier. In this way， the method effectively solves the problem that traditional classification methods face with confusion or ambiguity at the boundary of complex ground objects. A spatial mapping method based on precise matching of geographical coordinates is also proposed. A mapping relationship is established between optical image pixels and grid elements of a digital surface model （DSM）. This mapping enables a high-precision spatial alignment between the numbered optical land cover classification map and the DSM. As a result， each grid element of the DSM can be accurately associated with a type number for the corresponding ground object. Then， corresponding dielectric constants are assigned to each number. Thus， a physically consistent electromagnetic environment is constructed according to the electromagnetic characteristics of different ground objects. Finally， the shooting and bouncing rays （SBR） algorithm is used to simulate the propagation of radar echoes， and a high-fidelity SAR image with the characteristics of spatially inhomogeneous media is generated. The results of numerical experiments show that this method can efficiently generate high-fidelity SAR images and provide effective support for large-scale SAR simulation and remote sensing applications.