A novel predictive framework for water quality assessment based on socio-economic indicators and water leaving reflectance

Abstract

Accurate prediction of water quality at both spatial and temporal scales for large water bodies remains a daunting task with significant implications for human well-being and sustainable development (aligned with SDG 6 - clean water and sanitation). Traditional data-driven models on water quality prediction relied on some degree of field subsistence, which are neither cost-effective nor time-efficient. Socio-economic indicators have been concurrently used as predictor variable for water quality; however, such datasets typically available at coarse temporal resolutions, limiting their applicability for time-sensitive analyses. In this study, we integrated machine learning (ML) models with socio-economic indicators and remote sensing reflectance (R_RS) to address the challenge of temporality in predicting Biochemical Oxygen Demand (BOD) and Total Coliform Bacteria (TCB) levels across 228 lake systems in the Indian subcontinent. Pearson correlation analysis revealed limited direct correlations (<0.5) between BOD, TCB, and the input variables. However, a stepwise omission and commission analysis demonstrated that incorporating R_RS into the socio-economic model significantly enhanced predictive performance of the ML models. This approach achieved high classification accuracy for BOD and TCB, with Area Under the Curve (AUC) scores of 0.84 and 0.96, respectively, highlighting strong potential for temporal water quality assessment. Among the supervised learning methods tested, the random forest model outperformed all others in terms of accuracy and robustness. This study presents an integrated framework capable of predicting BOD and TCB with both high temporal and spatial resolution, and offers valuable insights for the effective and efficient management of aquatic ecosystems, enabling timely interventions and informed decision-making.