Sandy desertification poses a severe threat to global ecosystems, particularly in the Belt and Road Initiative (BRI) drylands, necessitating accurate mapping of sand dune/sheet (SDS) dynamics for sustainable land management. However, large-scale multi-temporal SDS mapping remains challenging because of the limited high-quality samples and spectral domain shifts across cross-temporal images. This study therefore proposed a novel deep learning domain adaptation approach for cross-temporal SDS mapping. Firstly, we designed a label self-filtering strategy to extract source-domain samples. The Fourier-based spectral adaptation was then introduced to mitigate cross-temporal domain shifts, followed by a TransUnet-integrated self-training method to refine predictions on unlabeled target years. Finally, a posterior optimization strategy was developed to correct inconsistent classifications. This approach successfully generated 30 m SDS maps for BRI drylands in 2000, 2005, 2010, and 2015, achieving an mIoU of 76.5%-83.8%. Spatiotemporal analysis revealed a total SDS area reduction of 4.735 & times; 10(4) km & sup2; over the 15-year period. Regionally, West Asia and Egypt witnessed SDS area expansion dominated by a shifting SDS, whereas East Asia, Central Asia, and South Asia experienced SDS reduction. This study provides a technical framework for large-scale cross-temporal SDS classification, and the generated SDS maps offer essential data support for desertification assessment in the BRI region.