Uncertainty-Aware Geospatial Foundation Model for Delineating Wetlands, Open Water, and Terrestrial Boundaries in a Tropical Lake Basin System

Wetlands in tropical lake basins are hydrologically dynamic transition zones that challenge conventional Earth observation classification owing to spectral ambiguity, seasonal inundation, and sensor-specific inconsistencies. This study proposes an uncertainty-aware hybrid framework coupling pretrained geospatial foundation model embeddings with a physically informed two-stage Random Forest classifier for robust delineation of wetland–water–land boundaries in the Lake Victoria Basin, Kenya. Multi-sensor Earth observation data, Sentinel-2 multispectral imagery and Sentinel-1 C-band SAR (VV/VH) are integrated at the feature level through a cross-season stacking architecture. Pretrained OLMoEarth V1 Nano embeddings (128-dim) extracted from 32×32-pixel chips are fused with per-chip spectral texture metrics (WIW, MNDWI) and SAR backscatter statistics. An updated Wetland Uncertainty Field (WUF)., combining predictive entropy, temporal instability, and sensor disagreement, each robustly normalized and equally weighted, characterizes classification uncertainty across hydrological gradients. The framework achieves consistent cross-season accuracy (overall accuracy ≈ 90.8%, Cohen’s kappa ≈ 0.86) for wet (March–May) and dry (June–September) seasons, with wetland F1-scores of 0.856 and 0.850. WUF validation via Mann–Whitney U tests show false-negative wetland errors carry significantly higher uncertainty than true negatives (p < 1×10⁻³⁰⁰ in both seasons), confirming that the uncertainty surface is spatially structured and environmentally meaningful. A seasonal dynamics product distinguishes permanent wetlands (33,274 ha) from seasonal wetlands (2,904 ha) and reverse-seasonal transitions (297 ha). This study advances geospatial foundation model applications by transitioning from deterministic classification toward probabilistic environmental representation, with implications for wetland monitoring, flood risk assessment, and climate adaptation planning in data-limited tropical environment