Integrating an interpolation technique and AI models using Bayesian model averaging to enhance groundwater level monitoring.

Monitoring groundwater levels in areas experiencing depletion is crucial for effective resource management. This study combines two approaches for estimating groundwater levels in regions lacking sufficient data for better spatial distribution estimates. To achieve this, several Artificial Intelligence (AI) models with different input features were developed using monthly groundwater level data from 2010 to 2023 in the Sacramento Valley, California. The results indicated that the Random Forest (RF) and Gradient Boosting Regressor (GBR) models, with Root Mean Square Error (RMSE) of 7.03 m and 7.83 m in the testing phase, respectively, were the most accurate. Subsequently, the data for each year in 2010–2023 were interpolated using the Ordinary Kriging (OK) method. The outputs of this method and the outputs from RF and GBR models were then merged using Bayesian Model Averaging (BMA). For 2010, 2015, 2020, and 2023, this approach reduced groundwater level estimation errors by 31.18, 41.87, 50.60, and 45.04%, respectively. Additionally, the results showed that the integrating method could reduce groundwater level estimates’ RMSE and Mean Absolute Error (MAE) by an average of 41.12 and 33.72% over 2010–2023. [ABSTRACT FROM AUTHOR]