The invention discloses a method for generating subtitles of a Transformer remote sensing image based on theme guidance, which mainly solves the problems that the generated description is single and semantic information in the image cannot be accurately represented in the prior art Its implementation scheme is as follows : a topic encoder composed of Transformer and topic vector is built, and pre-trained on the classified data set; build a semantic decoder composed of random mask layer, embedding layer, Transformer decoder and soft ‑ max layer; connect the theme encoder and the semantic decoder to obtain a remote sensing image subtitle generation network; setting training parameters and iteratively training remote sensing image caption generating network with standard RSICD data set; generate subtitle descriptions using the trained remote sensing image subtitle generation network. The invention improves the diversity and accuracy of the generated description, and can be used for ground object image retrieval, disaster prediction, and image understanding.