Data-driven water quality mapping of urban water bodies for sustainable cities and communities: integrating smartphone sensing and geographic information systems—a case study of Beira Lake, Sri Lanka

Effective management of degraded urban water bodies requires spatially dense, low cost, timely water quality monitoring and decision making system. Yet conventional laboratory monitoring remains too costly and time consuming for most cities. This study develops and validates a Geospatial Artificial Intelligence (GeoAI) framework integrating consumer smartphone multi-color-space imaging with Random Forest regression, SHapley Additive exPlanations (SHAP) explainability, Random Forest Spatial Interpolation (RFSI), and a governance centered Decision Error Analysis for urban lake water quality mapping. The framework is applied at Gangarama Lake, the Southwest Basin of Beira Lake, Colombo, Sri Lanka. Past studies have shown that it is the most severely polluted of the four Beira Lake basins (Water Pollution Index 1.02–3.94; COD up to 373.41 mg/L). Data were collected through direct site visits, with the dataset for this study consisting of smartphone images — multiple images (more than six) captured at each of 15 georeferenced locations. Colour features extracted across five colour spaces from images collected using an iPhone 14 Pro were processed through a five-stage GeoAI pipeline: Pearson screening, feature selection, Multiple Linear Regression (MLR) baseline, Random Forest ensemble regression with 5-fold cross-validated hyperparameter optimisation, and SHAP analysis. The primary applicability evaluation uses a Decision Error Analysis, assessing whether predictions lead managers to the correct threshold-based action rather than Root Mean Square Error (RMSE), which is dimensionally specific and scale-dependent, and thus not transferable across study sites. Across 60 threshold assessments (4 parameters × 15 locations), the framework achieved 97% correct governance decisions, zero missed violations, and two conservative false alarms at borderline locations. Outside the ±10% threshold boundary zone, decision accuracy was 100%. The Random Forest multi-color-space model achieved R 2 = 0.84 for turbidity and 0.81 for total phosphorus, outperforming single-space MLR by 8–22 R 2 percentage points. RFSI outperformed traditional Inverse Distance Weighted (IDW) and Kriging by 14–35% in Leave-One-Out Cross-Validation (LOOCV) RMSE through incorporation of distance-to-inlet spatial covariates. The framework reduces monitoring cost from conventional laboratory-only analysis to smartphone-based analysis.