Accurate and efficient soybean yield prediction is essential for ensuring food security and optimizing agricultural management. Remote-sensing approaches based on a single phenological stage or sensor may not fully capture the physiological and structural dynamics associated with soybean yield formation, thereby constraining prediction accuracy under field-level spatial heterogeneity. This study therefore developed and evaluated an integrated multi-source and multi-temporal UAV-based framework for plot-scale soybean yield prediction under production-field conditions. A UAV platform equipped with multispectral, RGB, and LiDAR sensors was used to extract multidimensional remote-sensing features at four key growth stages: beginning pod, full seed, beginning maturity, and full maturity. By integrating spectral indices, texture metrics, and three-dimensional canopy structural parameters, seven machine-learning algorithms—Ridge, LASSO, Random Forest, MLP, LightGBM, XGBoost, and CatBoost—were evaluated under single-temporal and multi-temporal scenarios. Nonlinear tree-based ensemble models generally achieved higher predictive accuracy than the linear models. On the independent test set, CatBoost performed best under the full multi-temporal, three-sensor fusion scenario, with an R² of 0.816 and an RMSE of 245 kg ha⁻¹. Among individual growth stages, the full seed stage (R6) produced the highest single-stage prediction accuracy. Three-sensor fusion improved prediction relative to the single-source multispectral scheme, and integrating features across phenological stages further improved performance by representing cumulative crop-growth dynamics. The proposed framework provides methodological support for UAV-based precision management and data-driven decision-making in soybean production.