A dynamic scene generation model for continuous observation of rocket in boost phase by infrared sensor in absorption band of satellite was established. A method of generating high-resolution surface emissivity images at band 22 and 23 of MODIS data based on neural networks was proposed. According to the proposed method, surface emissivity images with resolution of 100 meters were generated. Spectral emissivity of 4.18-4.5 μm was calculated by spectral correlation method; flight trajectories of rocket in boost phase were generated by Runge-Kutta method, and plume radiation transmission was calculated by LOS method to generate rocket plume image. The geometric relationship among rocket plume, surface points and the sensor on satellite was established. The plume and background were projected and imaged, and the dynamic scene of the rocket plume observed by the satellite was synthesized. By analyzing the radiance image sequences, it was found that radiances of the ground background was suppressed. At the same time, the target radiance contrast and the number of pixels occupied at different times were analyzed combined with the trajectory data. Furthermore, the difference of total radiation intensity curve of plume in different scenes was analyzed. The results show that the scene generation method is accurate and reliable, which can provide data basis and target characteristics support for target detection and tracking research based on images observed by satellite.