Abstract: To address the negative impact of color bias in multi-temporal remote sensing images on the accuracy of urban construction site change detection, an Adaptive Color Transfer Cycle-Consistent Adversarial Network (ACT-CycleGAN) is proposed. In the proposed framework, a Dynamic Feature Mapping Module (DFMM) is embedded within a dual-input U-Net generator to extract color-style guidance features, enabling adaptive color transfer to enhance cross-temporal color consistency. Furthermore, a content loss function is designed to constrain multi-level features, thereby preserving the structural and textural stability of ground objects, while an additional perceptual loss mitigates color distortion caused by strict content constraints. Experimental results demonstrate that, compared with existing mainstream color transfer algorithms, the proposed method achieves 4.62% improvement in SSIM and 2.73% in FSIM, while reducing ΔE₀₀ and FID by 0.1158 and 3.3832, respectively. Finally, in change detection experiments targeting urban construction site scenes, all evaluation metrics shows improvement after image transfer processing. This method effectively corrects color discrepancies between multi-temporal remote sensing images, thereby harmonizing the tonal characteristics of the inputs and reducing both false alarms and missed detections in change detection tasks. The results verify its high practical value for change monitoring in urban construction sites.