Rural roads extracted from agricultural machinery trajectories have significant research value. Due to the intricate network of rural roads and the large difference in the density of agricultural machinery trajectories, traditional road extraction methods struggle to perform effectively on rural roads with complex topology and on agricultural machinery trajectories with obscure road features. Therefore, this paper proposes a CGAN-based model named T2R-GAN (Trajectory to Road-GAN) for rural thematic road extraction, which learns the trajectory-to-road feature mapping through continuous adversarial training between the ELAU-Net generator and the PatchGAN discriminator to adapt to trajectories of various densities. ELAU-Net is an efficient network that utilizes an encoder-decoder structure and ELA modules to enhance the capture of obscure road features between sparse trajectories. To enhance model performance and reduce the risk of overfitting, bilateral hinge loss is designed to train our model to enhances the discriminator’s discriminative ability to facilitate the generator to generate more realistic roads improve the generalization of the model. To verify the effectiveness of T2R-GAN in extracting roads from agricultural trajectory data, this paper selects the real agricultural trajectory data from Henan Province, China in June 2021 as the dataset for verification. The experimental results show that the proposed method achieves 79.23% \(F1_{score}\) , which is 5.53% higher than the previous state-of-art. The proposed T2R-GAN provides a novel and effective approach for extracting rural thematic roads from agricultural machinery trajectories for the first time.