Random forest estimation of benthic chlorophyll-a in an intertidal flat using Sentinel-2: Uncertainty, predictor stability, and temporal robustness

Chlorophyll-a (Chl-a) is a key proxy for microphytobenthos (MPB) biomass and an indicator of water quality. Despite the increasing use of Sentinel-2 imagery and Random Forest (RF) models for chlorophyll-a (Chl-a) estimation, there has been insufficient evaluation of complementary RF strategies, epistemic uncertainty, hyperparameter optimization, and the temporal stability of predictor importance. We addressed these gaps using field-measured Chl-a, 20 Sentinel-2-derived predictors and RF models. Field Chl-a showed pronounced spatial and seasonal variability. Hyperparameter tuning identified split-rule selection as the primary determinant of performance, with ExtraTrees consistently optimal, whereas further tuning produced only marginal improvements. Epistemic uncertainty across 100 model replicates was low (relative uncertainty = 1.65%; RMSE change = 0.112 μg/g), indicating high model stability and reproducibility. LODO cross-validation demonstrated good temporal transferability for typical Chl-a concentrations on unseen sampling dates, although accuracy declined at extreme concentrations. The winter (January) seasonal model achieved the best predictive performance (RMSE = 2.71 μg/g; MAE = 2.39 μg/g), whereas the summer (July) model performed worst, underperforming both global and temporal RF models. Across modeling approaches, Sentinel-2 B4 and the Normalized Difference Turbidity Index (NDTI) were consistently the most influential and temporally robust predictors. Spatial predictions differed among RF approaches primarily in magnitude rather than spatial pattern. This integrated, uncertainty-aware, and reproducible framework strengthens multi-temporal Chl-a prediction and supports ecological monitoring.