Water level monitoring based on water gauge has been widely used because of its low cost and portability.However, due to the lack of water-scale image data sets and the relatively low quality of currently known data sets, a context-adjusted GAN（CA-GAN）is proposed in this paper. The experimental results show that CA-GAN images are of high quality，rich in scale and variety, and short generation time. The generated image data can provide image data for water level gauge and other applications based on deep learning. This study designs a new downsampling scheme with semantic activation and modulation by using the segmentation graph and Unet++ structure based on downsampling module is adopted. Meanwile, a context adjustment scheme between adjacent layers is designed in the study. The results show that the proposed downsampling module and intensive context adjustment module are significantly helpful to maintain the semantic information of water gauge images. Compared with CoGAN, SimGAN and CycleGAN models，the water image generated by CA-GAN can improve the measurement accuracy of deep network. In addition, CA-GAN model improves the accuracy of the generated water gauge image by 17% per pixel．