Explainable artificial intelligence and model calibration for water quality prediction

Abstract

Water is a key necessity for survival and sustenance of all living creatures. In the past years, the quality of water has been adversely affected by pollutants and other harmful wastes. This increased water pollution deteriorates water quality, making it unfit for any type of use most especially compromising the safety of drinking water for public health. The ecological safety and human health have continuously lowered due to hazardous pollution factors like chemicals and pathogens. By monitoring the Water Quality data parameters and forecasting them to get early warning, we can manage the quality of the water for different water sources. Numerous innovative technologies are slowly replacing human labor and other state of the art methods in water quality evaluation. Recently, different machine learning and artificial intelligence techniques have been adopted for water quality modeling which has become very beneficial in assessment and management of water resources. However, they suffer many times from high computational complexity, high prediction error and the blackbox nature in which they remain. Another big challenge faced by policy makers and other responsible Public Health Authorities is the lack of a relatively generalizable model for water quality prediction for public consumption with provision of explanations for understanding the most influential water quality parameters. This work presents an Explainable Artificial Intelligence method, SHAP (SHapley Additive exPlanations) to transparently and explainably assess the most important metrics that these models use in determining water quality based on potability. We also model a robust generalizable calibrated ensemble machine learning model for water quality prediction based on water potability and other water quality metrics from various water quality samples around the world. We then implement Automated Machine Learning with Stacked Ensembling to compare its results with those achieved by the Soft Voting Ensemble Model. The simulated results will provide theoretical support to policy makers and would be of interest to water planners in terms of assessing or maintaining water quality and improving sustainable pollu tion control, water and ecological management plans of water resources as well as early risk assessment and prevention in water environment in a simple, fast and cost-effective way which will protect the health of the people.