Integrating Earth observation data and metaheuristic feature selection for groundwater level forecasting: a methodological validation in Prince Edward Island, Canada
7 affiliations
- 1Canadian Centre for Climate Change and Adaptation, University of Prince Edward Island, St Peter's Bay, PE, Canada
- 2School of Business, Law, Humanities, and Pathways, University of Southern Queensland, QLD 4350, Australia
- 3Canadian Centre for Climate Change and Adaptation, University of Prince Edward Island, St Peter's Bay, PE, Canada; Faculty of Sustainable Design Engineering, University of Prince Edward Island, Charlottetown, PE, Canada; Corresponding author at: Canadian Centre for Climate Change and Adaptation, University of Prince Edward Island, St Peters Bay, PE, Canada.
- 4Faculty of Sustainable Design Engineering, University of Prince Edward Island, Charlottetown, PE, Canada; School of Computer Science, University of Guelph, Guelph, ON, Canada
- 5Department of Earth and Atmospheric Sciences, University of Alberta, Edmonton, AB, Canada
- 6Department of Crop and Soil Science, College of Agricultural, Human and Natural Resource Sciences, Washington State University, Pullman, WA 99164, United States
- 7School of Computer Science, University of Guelph, Guelph, ON, Canada; Corresponding author.
Abstract
Groundwater is an important resource, especially in areas with limited surface water. Nevertheless, groundwater level (GWL) forecasting is difficult because of complicated hydrogeological processes and the unavailability of data. This study introduces a hybrid modeling approach that combines a feature selection procedure, random forest (RF), and the Runge-Kutta optimizer (RUN), along with machine learning models, using Earth observation (EO) data to enhance GWL forecasting in Prince Edward Island, Canada. The RF-RUN model can be applied to determine the most significant predictors among a large number of predictors in a dataset using advanced feature selection algorithms. The machine learning models, that is, histogram gradient boosting (HGB), Elastic net, AdaBoost, and LASSO, were compared under the same evaluation criteria as SelectKBest and RF-RUN feature selection. The findings demonstrate that the RF-RUN–HGB model performed the most balanced and robust forecasting framework under multi-metric evaluations. In particular, the model yielded R 2 and RMSE values of 0.8352 and 0.3727 m and 0.8935 and 0.6196 m at the Sleepy Hollow and Caledonia stations, respectively. SHapley Additive exPlanations (SHAP) were applied to guarantee physical hydrogeological consistency to ensure the reliability of the framework. The SHAP analysis showed that although antecedent groundwater level was the strongest predictor at both sites, the shallow Sleepy Hollow well was very sensitive to high surface-atmosphere fluxes, such as reference evapotranspiration and precipitation. Conversely, the deeper Caledonia well was forced by the processes of delayed mass percolation, which depended greatly on lagged snow water equivalent and surface runoff. The sequential time-series conformal prediction framework (EnbPI) demonstrated high operational reliability, providing overall horizon coverage rates of 99.1% (PI = 1.81 m) at Sleepy Hollow and 97% (PI = 3.05) at Caledonia.
Related
- Integrating high-resolution magnetic surveys and metaheuristic optimization for quantitative characterization of subsurface magnetic sources in a municipal landfill
- Integrating Sentinel-1 Radar Soil Moisture and Explainable Machine Learning to Quantify Irrigation-Induced Groundwater Fluxes in the Hajdúság Region, Eastern Hungary
- Monitoring and forecasting groundwater storage dynamics in arid Wadi systems using GRACE/GRACE-FO satellite observations and XGBoost machine learning: Wadi El-Assiuti, Egypt (2003–2030)
- A hybrid physics–ML framework for integrating groundwater dynamics into land surface modeling: a ParFlow-based demonstration in headwater watersheds of the Pearl River Basin, China
- Uncertainty-Aware End-to-End AI Weather Forecasting: Disentangling Observation and Model Contributions
BibTeX
@article{karbasi2026integrating,
title = {{Integrating Earth observation data and metaheuristic feature selection for groundwater level forecasting: a methodological validation in Prince Edward Island, Canada}},
author = {Masoud Karbasi and Mumtaz Ali and Aitazaz Ahsan Farooque and Mehdi Jamei and Hassan Afzaal and Monireh Faramarzi and Syed Hamid Hussain Shah and Gurjit S. Randhawa},
journal = {Ecological Informatics},
year = {2026},
month = nov,
doi = {10.1016/j.ecoinf.2026.104045},
url = {https://doi.org/10.1016/j.ecoinf.2026.104045},
note = {CC BY},
}
RIS
TY - JOUR
TI - Integrating Earth observation data and metaheuristic feature selection for groundwater level forecasting: a methodological validation in Prince Edward Island, Canada
AU - Masoud Karbasi
AU - Mumtaz Ali
AU - Aitazaz Ahsan Farooque
AU - Mehdi Jamei
AU - Hassan Afzaal
AU - Monireh Faramarzi
AU - Syed Hamid Hussain Shah
AU - Gurjit S. Randhawa
PY - 2026
DA - 2026/11/16
JO - Ecological Informatics
DO - 10.1016/j.ecoinf.2026.104045
UR - https://doi.org/10.1016/j.ecoinf.2026.104045
AB - Groundwater is an important resource, especially in areas with limited surface water. Nevertheless, groundwater level (GWL) forecasting is difficult because of complicated hydrogeological processes and the unavailability of data. This study introduces a hybrid modeling approach that combines a feature selection procedure, random forest (RF), and the Runge-Kutta optimizer (RUN), along with machine learning models, using Earth observation (EO) data to enhance GWL forecasting in Prince Edward Island, Canada. The RF-RUN model can be applied to determine the most significant predictors among a large number of predictors in a dataset using advanced feature selection algorithms. The machine learning models, that is, histogram gradient boosting (HGB), Elastic net, AdaBoost, and LASSO, were compared under the same evaluation criteria as SelectKBest and RF-RUN feature selection. The findings demonstrate that the RF-RUN–HGB model performed the most balanced and robust forecasting framework under multi-metric evaluations. In particular, the model yielded R 2 and RMSE values of 0.8352 and 0.3727 m and 0.8935 and 0.6196 m at the Sleepy Hollow and Caledonia stations, respectively. SHapley Additive exPlanations (SHAP) were applied to guarantee physical hydrogeological consistency to ensure the reliability of the framework. The SHAP analysis showed that although antecedent groundwater level was the strongest predictor at both sites, the shallow Sleepy Hollow well was very sensitive to high surface-atmosphere fluxes, such as reference evapotranspiration and precipitation. Conversely, the deeper Caledonia well was forced by the processes of delayed mass percolation, which depended greatly on lagged snow water equivalent and surface runoff. The sequential time-series conformal prediction framework (EnbPI) demonstrated high operational reliability, providing overall horizon coverage rates of 99.1% (PI = 1.81 m) at Sleepy Hollow and 97% (PI = 3.05) at Caledonia.
SN - 1574-9541
SN - 1878-0512
ER -