Gaussian Process Regression for Spatio-Temporal Reconstruction and Uncertainty Quantification of Satellite-Derived Chlorophyll-a in Laguna de Bay, Philippines

Satellite-derived chlorophyll-a time series in tropical inland water bodies are frequently compromised by cloud cover, atmospheric interference, and sensor artifacts, producing spatially irregular and temporally inconsistent observation panels. This study evaluates Gaussian Process Regression (GPR) as a reconstruction framework for gap-affected chlorophyll-a estimates derived from Sentinel-3 OLCI imagery over Laguna de Bay, Philippines, spanning January 2021 to December 2025. A Modified Maximum Chlorophyll Index (MCI*), substituting the anomalous Oa10 band with Oa09, was computed across 895 one-kilometer grid cells, yielding a panel of 53,700 cell-month observations with 1.42% missingness following quality filtering and outlier removal. Monthly-batch GPR with a composite kernel combining a scaled Radial Basis Function and a scaled Matérn 3/2 component was fitted independently per month, producing posterior mean reconstructions alongside calibrated uncertainty estimates. Reconstruction accuracy was benchmarked against Inverse Distance Weighting (IDW) and Ordinary Kriging (OK) using leave-one-out cross-validation evaluated on RMSE, MAE, and R². GPR achieved an aggregate R² of 0.695, outperforming OK (R² = 0.648) and IDW (R² = 0.562), and won 44 of 60 monthly folds. Beyond accuracy, GPR uniquely provides posterior predictive uncertainty maps that expose spatially heterogeneous reconstruction reliability—a capability unavailable from deterministic or variogram-based interpolants. These findings position GPR as a principled and scalable reconstruction method for satellite-derived aquatic variables in optically complex, data-sparse tropical environments.