Abstract: Radar image interpretation is one of the key technologies for promoting wide applications of radar satellites and supporting future unmanned intelligent platforms. Microwave vision, as a perceptual inverse problem for understanding the physical world through electromagnetic data, focuses on deriving semantic information from microwave radar images according to physics principles. Its forward problem involves "microwave graphics" that characterize the interaction mechanisms between electromagnetic waves and the physical world, aiming to develop electromagnetic scattering modeling suitable for perceptual inverse problems. This paper proposes a semantic electromagnetic scattering modeling framework for radar intelligent perception, which centers on target semantics and introduces diversity-oriented randomness modeling. This approach shifts the focus from pursuing one-to-one consistency of individual samples to ensuring distributional consistency across sample populations. This paper elaborates on the problem background, fundamental properties, and key tasks of semantic electromagnetic scattering modeling, presenting technical roadmaps at two levels: the primitive scatterer dictionary and the semantic representation tree. Finally, some relevant research progresses of the authors are briefly introduced.