Study region: The research was undertaken across three arid sites in Saudi Arabia: Bisha, Buraydah, and Duba. Study focus: This paper presents a novel meta-learning framework for accurate monthly estimation of actual evapotranspiration (AET) using multimodal satellite data to support strategic irrigation planning and seasonal water resource allocation. The two-stage architecture integrates nonlinear feature transformations directly into a meta-learning layer, creating a self-optimizing pipeline that resolves complex dynamics beyond the reach of conventional methods. The data-set, which spans from the years 2003-2024, contains eight TerraClimate factors (TCF) and four surface-reflectance indices (SRI). Feature selection and hyperparameter optimization were performed via three machine learning algorithms: Penalized Spline (P-spline), Gradient Boosting Regressor (GBR), and Partial Least Squares Regression (PLSR). A meta-learning approach upgraded the behavior of models for forecasting monthly AET with the best performing models being designated based on statistical agreement with observations. New hydrological insights for the region: The analysis results reveals that the proposed P-spline_Pspline two-stage modelling architecture comprising the deployment of higher-level variables generated by P-splines as inputs to a P-spline meta-learner yields superior predictive outcomes for AET. The integrated process produced an R2 of 0.923 and an RMSE of 5.337 mm, which was better than the model that employed only the P-spline with an R2 of 0.914 and an RMSE of 5.671 mm. This hybrid methodology leverages coupled P-spline feature engineering and meta-learning to estimate AET in hyper-arid environments, eliminating reliance on scarce in-situ measurements through satellite-data-driven nonlinear transformation. In turn, the optimized structure paves the way for more effective irrigation scheduling and sustainable water management practices.