Spatial prediction and mapping of water quality of Owabi reservoir from satellite imageries and machine learning models

Abstract

Estimation and mapping of surface water quality are vital for the planning and sustainable management of inland reservoirs. The study aimed at retrieving and mapping water quality parameters (WQPs) of Owabi Dam reservoir from Sentinel-2 (S2) and Landsat 8 (L8) satellite data, using random forests (RF), support vector machines (SVM) and multiple linear regression (MLR) models. Water samples from 45 systematic plots were analysed for pH, turbidity, alkalinity, total dissolved solids and dissolved oxygen. The performances of all three models were compared in terms of adjusted coefficient of determination (R2.adj), and the root mean square error (RMSE) using repeated k-fold cross-validation procedure. To determine the status of water quality, pixel-level predictions were used to compute model-assisted estimates of WQPs and compared with reference values from the World Health Organization. Generally, all three models produced more accurate results for S2 compared to L8. On average, the inter-sensor relative efficiency showed that S2 outperformed L8 by 67% in retrieving WQPs of the Owabi Dam reservoir. S2 gave the highest accuracy for RF (R2.adj = 95–99%, RMSE = 0.02–3.03) and least for MLR (R2.adj = 55–91%, RMSE = 0.03–3.14). Compared to RF, SVM showed similar results for S2 but with slightly higher RMSEs (0.03–3.99). The estimated pH (7.06), total dissolved solids (39.19 mg/L) and alkalinity (179.60 mg/L) were within acceptable limits, except for turbidity (33.49 mg/L) which exceeded the reference thresholds. The S2 and RF models are recommended for the monitoring of surface water quality of the Owabi Dam reservoir.

Keywords

• Water quality
• Optical satellite image data
• Machine learning models
• Owabi Reservoir