Large-scale atmospheric circulations inferred from manually labeled dunes may be biased due to limitations in data quantity and regional representativeness. This study developed a post-processing workflow based on the adjacent dune similarity and applied it to mask region-based convolutional neural network (Mask R-CNN), extracting morphological metrics for 206,972 barchans (5113 with dynamics) in the Taklamakan Desert. Our results show that dune-derived near-surface wind circulations generally align with reanalysis data but exhibit significant deviations near mountain ranges due to topographic effects. Compared to static dune morphologies, multi-temporal analysis of dune migration provides a more reliable reference for identifying dominant sandtransporting winds by minimizing the effects of local wind variability, thus serving as a robust indicator for reconstructing long-term wind regimes. Dune dynamic patterns in the desert interior regions are more complex compared to those in the marginal areas. When using interior dunes to predict local circulation, it is essential to account for the influences of local topography and dune morphology variations. After correcting for the dune-size effect, the sand flux estimated from our dune celerity dataset demonstrates comparable magnitude and spatial distribution to reanalysis predictions. Notably, in areas with high barchan density, dune-based sand flux estimation reveals fine-scale variations in aeolian sediment transport.