Remote sensing image change detection is widely applied in urban planning, environmental monitoring, and disaster assessment. However, satellite remote sensing images typically exhibit complex scene characteristics and irregular texture patterns, leading to frequent false changes and missed detection of fine-grained targets when deep learning methods are used for change detection. To address these issues, a remote sensing image change detection method based on an improved deep convolutional generative adversarial network was proposed. A U-Net generator combined with coordinate attention was designed to enhance the model's sensitivity and accuracy in detecting fine-grained targets under complex scenes. Simultaneously, an improved lightweight fully convolutional network was adopted as the discriminator, effectively resolving problems of incomplete or missing detection features through generative adversarial training. Furthermore, a weighted loss function was developed to balance model training stability and detection accuracy. Experiments were conducted on the public change detection datasets SYSU-CD (Sun Yat-Sen University-change detection) and CDD (change detection dataset) demonstrated that the proposed method achieved significant performance improvements. The method achieves improvements of 3.02% and 5.76% in F₁ scores compared to the baseline UNet, reaching final scores of 80.59% and 96.25%, respectively, which are superior to other comparative methods.