Multisensor Ensemble Mapping of Sub-hectare Ephemeral Surface Water in Kenyan ASALs

Ephemeral pans and seasonal ponds in arid and semi-arid lands supply critical water for pastoral and ecological systems, yet are not routinely monitored due to their small size, highly dynamic and spectral confusion with vegetation and shadows. We present and evaluate a multisensor mapping approach to detect sub-0.5 ha surface water bodies and quantify their linkage to rainfall variability to inform decision making.

Our approach fuses Sentinel-1 SAR, Sentinel-2 optical indices and DEM derived covariates within an ensemble classifier (voting of Random Forest, Gradient Boosting, and Decision Tree models). Predictive uncertainty is mapped using ensemble agreement and class probabilities, and we compare SAR-only, optical-only, terrain-only, and fused configurations. Additionally, rain and ephemeral surface water dynamics are modelled using generalised additive models with CHIRPs  and local rain gauge observations to test the lagged relationships in monthly water area anomalies.

Results show the fused model achieves an overall accuracy of 85%, outperforming Sentinel-1, and Sentinel-2 (78% and 72%, respectively). Generalised additive models explain 62% of variance in monthly water area anomalies, with a strong response at 1-3 month lags. These results show multisensor fusion with  quantified uncertainty improves detection of ephemeral surface water and enables estimation of rainfall thresholds and lagged dynamics relevant to pastoral water planning and targeted anticipatory action interventions.