Combining high resolution input and stacking ensemble machine learning algorithms for developing robust groundwater potentiality models in Bisha watershed, Saudi Arabia

Abstract

Abstract The present research aims to build a unique ensemble model based on a high-resolution groundwater potentiality model (GPM) by merging the random forest (RF) meta classifier-based stacking ensemble machine learning method with high-resolution groundwater conditioning factors in the Bisha watershed, Saudi Arabia. Using high-resolution satellite images and other secondary sources, twenty-one parameters were derived in this study. SVM, ANN, and LR meta-classifiers were used to create the new stacking ensemble machine learning method. RF meta classifiers were used to create the new stacking ensemble machine learning algorithm. Each of these three models was compared to the ensemble model separately. The GPMs were then confirmed using ROC curves, such as the empirical ROC and the binormal ROC, both parametric and non-parametric. Sensitivity analyses of GPM parameters were carried out using an RF-based approach. Predictions were made using six hybrid algorithms and a new hybrid model for the very high (1835–2149 km2) and high groundwater potential (3335–4585 km2) regions. The stacking model (ROCe-AUC: 0.856; ROCb-AUC: 0.921) beat other models based on ROC's area under the curve (AUC). GPM sensitivity study indicated that NDMI, NDVI, slope, distance to water bodies, and flow accumulation were the most sensitive parameters. This work will aid in improving the effectiveness of GPMs in developing sustainable groundwater management plans by utilizing DEM-derived parameters.

Keywords

• Data mining
• Hybrid ensemble machine learning
• Logistic regression
• Sensitivity analysis
• Remote sensing