The detection of multitemporal semantic changes is often used to monitor changes in agricultural ecology and to track the development of agricultural land because it uses semantic information to analyze the specific types of changes. A method of detecting multitemporal semantic changes using weak labels with noise and low resolution instead of high-resolution labels was proposed to solve the problem that the scarcity of high-resolution remote sensing image labels and the slow growth of labeling technology limit the development of detecting multitemporal semantic changes. First，low resolution satellite data were used to smooth the quality differences of high-resolution remote sensing image inputs. Secondly，the high-resolution remote sensing image classification map was estimated by combining the epitomes model and the label super-resolution algorithm as a statistical inference algorithm，and a small FCN network was fitted to post-process the remote sensing image classification map generated to improve its classification. Finally，the results of detecting change were obtained by comparing the differences between different simultaneous land cover classification images. The results showed that the proposed method improved the mean intersection over uion （mIoU） by 8.9 percentage points compared with other methods of detecting multitemporal semantic changes，and detected the changes of land cover classification effectively.