Timely and accurate acquisition of root-zone soil moisture (SM) is critical for agricultural precision irrigation. This study combined UAV multispectral data with three machine learning algorithms (MLAs)—back-propagation neural network (BP), random forest (RF), and extreme random tree (ET)—to invert maize root-zone SM across different growth stages and water stresses. Results indicated that sensitive vegetation indices (VIs) at various growth stages were mainly related to NIR and R bands. For different growth stages, the optimal inversion depth for all three models was 0–45 cm, and the RF model performed best, especially during the grain filling–maturation stage (R² = 0.794, RMSE = 2.043%). Under different water stresses, the RF model achieved the optimal performance under the mild stress irrigation treatment, especially for T2 (the irrigation upper limit of 90%), with an R2 of 0.827 and an RMSE of 2.593%. UAV multispectral data combined with MLAs can accurately estimate maize root-zone SM, supporting scientific farmland irrigation.
Guo et al. (2026) studied this question.