This study investigates the spatiotemporal evolution of carbon sources/sinks in arid urban ecosystems through Net Ecosystem Productivity (NEP) analysis, aiming to reconcile urban expansion with ecological conservation in Urumqi, a core city on China's Silk Road Economic Belt. Combining multi-source remote sensing data (MODIS, Landsat, NPP-VIIRS) with socioeconomic datasets, we developed a hybrid framework integrating CASA model and soil respiration algorithms to quantify NEP dynamics in Urumqi City (2005-2020). Machine learning approaches (Random Forest) and spatial econometric models were applied to identify dominant drivers, with particular focus on policy-induced land use/cover change (LUCC). Three key findings emerge: (1) NEP exhibited 15-year cumulative growth (+ 27.3%), with carbon sink hotspots concentrating in ecologically restored southern suburbs (910.14 g C m(-)& sup2; yr(-)& sup1;), contrasting with carbon source clusters in northern industrial zones (-19.68 g C m(-)& sup2; yr(-)& sup1;); (2) Following the 2010 ecological redline policies, localized improvements in carbon sequestration capacity were detected (average enhancement approximately 18.9% in high-response zones), although the overall NEP trend remained statistically stable across most of the study area (non-significant in > 95% of pixels). These results indicate that policy-induced LUCC facilitated spatially concentrated carbon sink strengthening rather than a citywide enhancement. (3) Random Forest modeling revealed LUCC as the predominant driver (18.36% importance), outweighing climate factors (precipitation: 12.7%, temperature: 9.4%) and socioeconomic parameters (NSL < 5%). Our findings challenge the urbanization-carbon loss paradigm by demonstrating targeted land use optimization as an effective policy instrument for dryland cities. The machine learning-enhanced framework provides transferable methodology for SDG 11 (Sustainable Cities) monitoring in arid regions.