Seagrass meadows are critical coastal ecosystems that provide invaluable services: carbon sequestration, habitat for marine biodiversity, and shoreline protection, but they are declining at alarming rates. Recent studies estimate arround 29% global loss of seagrass cover since the 1700s and project large carbon emissions if this loss continues. Conventional monitoring methods struggle to capture these complex dynamics, motivating data-driven approaches. In this work, we develop an integrated machine learning framework for spatio-temporal modeling of seagrass health and biomass. We apply iterative imputation to handle missing environmental data, and train both an XGBoost regressor and a stacked ensemble of ML models (Ridge, Random Forest, LGBM) on geospatial features. Model hyperparameters are tuned via the Tree-structured Parzen Estimator (TPE), a Bayesian optimization technique that efficiently explores the hyperparameters. Our results show that the optimized XGBoost model outperforms the ensemble across key seagrass metrics, achieving superior accuracy in predicting biomass and productivity. We identify the most influential environmental drivers such as light availability and water temperature through feature importance analysis, providing actionable ecological insights. This enhanced predictive framework, supported by a clear methodology flow, offers a robust tool for proactive seagrass conservation and decision-making.