Long-term monitoring of aboveground biomass (AGB) is essential for assessing ecosystem productivity and carbon storage. However, sparse vegetation and complex bare-soil backgrounds in desert regions introduce significant uncertainties in AGB inversion and driver analysis. In this study, field plots, UAVs, and satellite data were integrated to develop high-accuracy AGB inversion models for desert ecosystems using vegetation classification, data fusion, and multi-model comparison. Theil-Sen slope estimation, Mann-Kendall tests, and SHAP analysis were applied to characterize AGB dynamics and identify the driving mechanisms. The results show that object-based classification combined with deep learning (SegFormer) achieved the highest accuracy (OA = 94%). Vegetation-class-based variables improved UAV-scale AGB estimation. At the satellite-scale, spectral features generated by Landsat-MODIS fusion, combined with environmental variables, yielded the best performance with the Boruta-XGBoost model (R2 = 0.84 RMSE = 733.32 kg ha-1, CCC = 0.91). From 2000 to 2024, AGB exhibited an increasing trend, with 57.6% of the area increasing mainly in low-elevation temperate desert regions, while 37.9% decreased primarily in high-elevation temperate desert steppe regions. Temperature and precipitation were the dominant drivers of AGB variation, with notable interaction effects involving temperature-precipitation and potential evapotranspiration-precipitation. When precipitation exceeded 82 mm, the effects of temperature and potential evapotranspiration on AGB were enhanced and became positive influences. This study provides methodological support for AGB modeling and informs desertification management in arid regions.