This study aims to develop a high-accuracy root-zone SMC (Soil Moisture Content) inversion method suitable for cotton fields in arid regions, in order to support precision irrigation and agricultural water resource management. Field experiments were conducted from 2022 to 2025 in Tumushuke and Alaer, Xinjiang, covering multiple growth stages and regions. UAV-based RGB and thermal infrared imagery were combined with ground measurements of SMC, electrical conductivity, leaf water potential, and meteorological observations to build an integrated air-ground-atmosphere multi-source data fusion framework. A Bayesian-optimized XGBoost model was employed to predict root-zone SMC. Results showed that single-source data (referring to a single type of data, e.g., meteorological, in situ, or UAV remote sensing data) provided limited predictive accuracy, with driving data yielding an R2 of only 0.37 and RMSE of 10.81; in situ data performed relatively better, with R2 = 0.64 and RMSE = 6.80. Dual-source fusion (combining two types of data, e.g., UAV remote sensing and in situ data) significantly improved performance, with UAV and in situ data combined achieving R2 = 0.69 and RMSE = 6.06. The best results were obtained with tri-source fusion (integrating UAV remote sensing, in situ, and meteorological data), where R2 increased to 0.73 and RMSE decreased to 5.50. Multi-stage testing further demonstrated robustness: at the flowering stage, test sets in Alaer and Tumushuke reached R2 values of 0.75 and 0.76 with RMSE of 4.95 and 5.98, respectively; at the full boll stage, both sites achieved R2 = 0.70 with RMSE of 4.90 and 5.09; at the open boll stage, R2 values were 0.63 and 0.66 with RMSE of 4.97 and 5.95, respectively. These findings indicate that models integrating UAV, meteorological, and physiological multi-source data can effectively capture the spatiotemporal heterogeneity of SMC in arid cotton fields, providing a scientific basis for precision irrigation and agricultural water resource management.