Understanding vegetation dynamics in semi-arid regions is vital for ecological restoration. However, previous studies have been constrained by the lack of a continuous, high-spatial-resolution normalized difference vegetation index (NDVI) dataset around the year 2000, and by downscaling methods that rely excessively on image texture without sufficient constraints from environmental feature variables. Moreover, the key regulatory role of groundwater on vegetation in semi-arid areas has often been overlooked, leading to a biased understanding of vegetation cover dynamics and reduced reliability of conclusions. In this study, the Songnen Plain in Jilin Province, China, was selected as the research area. A synergistic downscaling approach combining random forest and nonlinear residual correction was developed to analyze the spatiotemporal dynamics of annual mean NDVI from 1985 to 2022. Driving factors of vegetation dynamics were analyzed using the geographical detector. The results indicate that the proposed downscaling approach achieves a higher fitting accuracy than conventional methods (R2 = 0.90). The regional NDVI exhibited a significant upward trend (growth rate = 0.001/a). Throughout the entire study period, the factor with the strongest explanatory power for NDVI in the Low Plain was LUCC, in the Piedmont Plain it was precipitation, and in the High Plain it was population density, with their respective multi-year averages being 0.22, 0.17, and 0.14. Although groundwater depth alone showed a low explanatory power, its interactions with climate, topography, and anthropogenic factors were notable. The developed downscaling method enables refined characterization of regional vegetation spatial patterns, providing a solid data foundation for ecological management. The findings enhance understanding of vegetation dynamics and provide guidance for regional ecological management.