Enhancing Groundwater Recharge Prediction: A Feature Selection-Based Deep Forest Model With Bayesian Optimisation

Accurate prediction of groundwater recharge is crucial for the sustainable management of water resources. Existing models, while effective, still have potential for improved accuracy. This study proposed a novel deep forest model-the Feature Selection-based Deep Forest model (FSDF)-for enhanced groundwater recharge prediction. This model consists of three key essential components: a feature selection layer, a cascade enhancement layer and a decision output layer, all designed to enhance the prediction accuracy of groundwater recharge rates. The feature selection layer effectively filtered out redundant features, ensuring that only relevant features are fed into the subsequent cascade enhancement layer. The cascaded enhancement layer was jointly constructed by random forests and completely random forests, processing the data layer-by-layer. Finally, the predictions of groundwater recharge rates were produced through an averaging strategy in the decision output layer. To further enhance the FSDF model's predictive capabilities, Bayesian optimization was applied for fine-tuning model hyperparameters. The model's performance was evaluated and compared with existing models using a dataset comprising of groundwater recharge rates from 1549 wells in New South Wales, Australia. The FSDF model exhibited exceptional performance, achieving a training accuracy of 95.91% and a testing accuracy of 89.65%. It outperformed the adaptive boosting, categorical boosting, extreme gradient boosting, multiple linear regression and random forests by 2.02%, 6.98%, 9.05%, 17.02% and 2.74% in prediction performance, respectively. This study contributes to both hydrological processes and groundwater management by identifying key factors such as rainfall, surface geology and PET, and refining hydrological models for greater predictive accuracy. The FSDF model offers a powerful tool for accurately forecasting groundwater recharge, outperforming traditional models. The model's adaptability makes it applicable to different geographical regions for managing water resources in the face of challenges such as water scarcity and climate change.