Integrated ensemble learning approach for multi-depth water quality estimation in reservoir environments

Water quality is paramount for the well-being of ecosystems and organisms, so assessing water quality variables (WQVs) is imperative. Despite existing research on predicting WQVs using machine learning models, none have tackled predicting WQVs at varying depths using surface water data. This study addresses this gap by collecting pertinent data and constructing an appropriate dataset. Utilizing AAQ-RINKO, a water quality profiling instrument, data was gathered from the Wadi Dayqah Dam reservoir in Oman. Leveraging surface water data (salinity, density, and temperature), the study estimates three WQVs - Dissolved Oxygen, Chlorophyll-a, and Turbidity - across depths ranging from 1 meter to 35 meters. A comprehensive evaluation of various machine learning and deep learning methods for tabular data was conducted alongside the introduction of seven ensemble approaches, with one emerging as the most effective. This approach leverages clustering to separate sub-models' importance in the final ensemble, enhanced by Bayesian optimization for weighting. Notably, Random Forest and Extreme Gradient Boosting techniques demonstrated superior performance following the proposed approach. Additional analyses assessed the impact of various inputs, sampling depth, time, and location. Overall, the proposed method exhibited significant enhancements, yielding improvements ranging from 4 % to 16 % based on the error metric.