This study aimed to assess the potential of machine learning models applied to high spatial resolution images from UAVs for estimating the aboveground biomass (AGB) of forage grass cultivated in the Brazilian semiarid region. The fresh and dry AGB were determined in Cenchrus ciliare plots with an area of 0.04 m². Spectral data were obtained using a multispectral sensor (Red, Green, and NIR) mounted on a UAV, from which 45 vegetation indices were derived, in addition to a structural variable representing plant height (H95). Among these, H95, GDVI, GSAVI2, GSAVI, GOSAVI, GRDVI, and CTVI exhibited the strongest correlations with biomass. Following multicollinearity analysis, eight variables (R, G, NIR, H95, CVI, MCARI, RGR, and Norm G) were selected to train Random Forest (RF), Support Vector Machine (SVM), and XGBoost models. RF and XGBoost yielded the highest predictive performance, both achieving an R² of 0.80 for AGB—Fresh. Their superiority was maintained for AGB—Dry estimation, with R² values of 0.69 for XGBoost and 0.67 for RF. Although SVM produced higher estimation errors, it showed a satisfactory ability to capture variability, including extreme values. In modeling, the incorporation of plant height, combined with spectral data obtained from high spatial resolution imagery, makes AGB estimation models more reliable. The findings highlight the feasibility of integrating UAV-based remote sensing and machine learning algorithms for non-destructive biomass estimation in forage systems, with promising applications in pasture monitoring and agricultural land management in semi-arid environments.