Remote sensing of water presence and physicochemical parameters in water pans in semi-arid Kenya using Sentinel-2 imagery

Climate change, population growth, and water source pollution present significant challenges for water resource managers, who often face limited resources and scarce in-situ data in the arid and semi-arid regions of Sub-Saharan Africa. In Kenya, about 80% of the country’s land area falls within these arid and semi-arid regions. These areas support 36% of the human population and 70% of the livestock population, contributing roughly 50% of agricultural GDP and 15% of national GDP. Water pans, which are small, shallow reservoirs that fill with surface runoff, sustain local livelihoods by supplying water for livestock, wildlife, and small-scale agropastoral irrigation. However, water pans are becoming increasingly susceptible to climate change, highlighting the need for continuous monitoring to ensure water availability for communities. Remote sensing provides a valuable complement to traditional water monitoring techniques by offering spatially extensive and repeatable observations that enhance our understanding of water dynamics. In this context, our primary aims were to evaluate the potential of Sentinel-2 imagery for detecting water presence and assessing physicochemical parameters, specifically turbidity and specific conductivity, in water pans in Taita Taveta County, Kenya. We employed mid-point and receiver operating characteristic (ROC) curve-based threshold methods for water presence detection, alongside generalized additive models for estimating turbidity and specific conductivity. We achieved an F1 score greater than 95% for water presence detection using the Normalized Difference Moisture Index and the two thresholds. The B8A/B4 predictor for specific conductivity yielded a coefficient of determination (R²) of less than 0.5 with both standard and group leave-one-out cross-validation (LOOCV). In contrast, the B8/B4 and B8/B5 predictors for turbidity recorded R² values greater than 0.8 with standard LOOCV and greater than 0.6 with group LOOCV. Overall, this study demonstrates the potential of remote sensing-based approaches for water monitoring, even under conditions of limited data availability.