Accurate and spatially representative soil moisture information at high-resolution is important for environmental monitoring, water resource management and sustainable agriculture in heterogeneous and water-scarcity areas. One of the biggest challenges in South African earth observation datasets is that existing satellite-based soil moisture products are limited by coarse spatial resolution, while in situ measurements are costly and spatially scarce. This study presents a soil moisture retrieval framework that combines multispectral imagery with a stacked ensemble machine learning approach to estimate field-scale surface soil moisture in a smallholder irrigated maize farm. Unmanned Aerial Vehicle (UAV) images were acquired across three key crop phenological stages and a soil-line-based soil moisture index (SMI) derived from near-infrared and red reflectance was combined with spectral features to train multiple machine learning models. The models included linear regression, ridge regression, support vector machines, k-nearest neighbours, decision tree and random forest. A stacked ensemble model was developed to combine individual model capabilities and enhance predictive performance. Evaluation against in situ measurements showed that the ensemble model outperformed other models with a coefficient of determination value (R²) of 0.93, Root Mean Square Error (RMSE) of 0.96%, Mean Absolute Percentage Error (MAPE) of 1.90% and Explained Variance (EV) of 91.19%) compared to individual models and NDVI-based estimates. The spatial distribution maps revealed persistent moisture gradients reflecting irrigation patterns and field heterogeneity with highest soil moisture at mid-development and lower levels at maturity due to canopy cover. This approach demonstrates the potential of UAV-based ensemble modelling for precision water management and improved environmental monitoring in water-limited agricultural systems.