Relative humidity quantification using interpretable machine learning based- stacking approach: representative case study in Ethiopia

Relative humidity (RH) is among the water cycle’s important parameters and stochastic processes. Accurate estimation of RH is essential for numerous water resources management tasks. In this study, five ensemble machine-learning models including CatBoost, random forest (RF), AdaBoost, extreme gradient boost (XGB), and Gaussian processing regression (GPR) were utilized to estimate RH at Hawassa Lake catchment in Ethiopia. Ten years of meteorological data were used for model development. Afterward, two model stacking techniques such as weighted average stacking (WAS) and neural network-based stacking (NNS) were developed by stacking the outcomes of the base models. The accuracy of the predictive models was measured using Nash Sutcliffe Efficiency (NSE), root mean square error (RMSE), mean absolute error (MAE), determination coefficient (R²), and different graphical illustrations. To determine the influence of different weather parameters on RH estimation results, Shapley Additive exPlanations (SHAP) analysis was performed, and maximum temperature, sunshine hour, and evaporation were found to be the most dominant variables. Among the individual models, AdaBoost attained the most accurate RH estimation with RMSE = 4.927%, MAE = 3.776%, NSE = 0.868, and R² = 0.872 in the testing period. The NNS technique achieved the highest predictive performance with NSE = 0.944 and RMSE = 3.196%, improving the base model’s efficiency between 35.133% and 42.07% based on testing set RMSE values. Generally, the study result revealed the strength of the proposed base models and stacking methods in RH estimation in semi-humid climates.