Background: Arid wetlands like Iran's Bakhtegan face severe desiccation from climate change and human activities. Monitoring these ecosystems is hindered by insufficient in-situ data and complex non-linear dynamics, making high-precision monthly forecasting in ungauged basins a significant challenge for water engineering.

Objective: This study develops a robust monthly surface water extent (SWE) forecasting framework using a multi-scale hybrid Stacking Ensemble. The goal is to integrate satellite data with explainable AI (XAI) to ensure physical interpretability and provide a reliable tool for early warning systems.

Methods: A 10-year (2013 to 2023) dataset of Landsat 8/9, CHIRPS, and MODIS data was extracted via Google Earth Engine. After Savitzky-Golay denoising, a two-layer Stacking Ensemble integrating XGBoost, ExtraTrees, and Ridge Regression was implemented. Hydrological memory was explicitly modeled using 1-month and 12-month lags alongside 6-month cumulative water balances.

Results: The framework achieved high accuracy with a Nash-Sutcliffe Efficiency (NSE) of 0.81, R-squared of 0.81, and PBIAS of -3.55 percent. SHAP analysis identified 12-month area persistence and 6-month cumulative hydro-balance as the primary drivers. The model successfully isolated anthropogenic signals, with late-summer residuals indicating unmodeled irrigation withdrawals and dam impacts.

Conclusions: This research offers a scalable and interpretable soft computing tool for water management in data-scarce regions. By providing reliable forecasts, it supports decision makers in optimizing water allocation and preventing ecological collapse in endangered hypersaline wetlands globally.