Under the context of global climate change, the spatiotemporal evolution and driving mechanisms of dust weather have become critical scientific issues for environmental security and risk governance in arid regions. Southern Xinjiang, which encompasses major dust source areas such as the Taklimakan and Kumtag Deserts, is one of the most severely affected regions in China. A comprehensive understanding of the spatiotemporal patterns and driving factors of dust weather in this region is essential not only for elucidating its formation mechanisms but also for providing a scientific basis for targeted prevention and management strategies. Based on ground-based observations and reanalysis/remote sensing datasets from 2000 to 2023, this study combines spatiotemporal analysis with machine learning methods to systematically investigate the evolution characteristics of dust weather in a typical source-oasis composite zone of southern Xinjiang. Specifically, key meteorological drivers were identified, their responses under different land-surface conditions were examined, and cluster analysis was applied to classify dominant meteorological regimes associated with different dust-event types.The main findings are as follows: (1) Dust-weather frequency exhibits an overall increasing trend, with significant increases at multiple stations during 2012-2023, along with pronounced seasonality and spatial heterogeneity. (2) Floating dust, blowing dust, and dust storms are all strongly influenced by wind speed, air pressure, and temperature; however, their sensitivities to these factors differ markedly. (3) Meteorological driving effects vary across land-surface conditions, reflecting a typical surface-atmosphere interaction mechanism. (4) Four distinct meteorological regimes are identified by cluster analysis, showing clear differences in dust-event associations even at the same site.This study deepens the understanding of dust-weather formation mechanisms in arid regions and provides scientific support for environmental governance and risk management in southern Xinjiang and Central Asia. The proposed framework, integrating multi-source data with machine learning, offers methodological guidance for improving dust-weather early-warning systems and promoting sustainable environmental management.