From Statistical Assessment to Data-Driven Forecasting: Groundwater Level Analysis and Modelling in Southern Ontario

Abstract

Groundwater is a vital resource in Southern Ontario, supplying 90% of rural residents and around 200 municipalities with drinking water and irrigation for agriculture. Accurate groundwater level (GWL) estimation is important for sustainable water management and environmental conservation. However, direct GWL measurements are often costly and spatially limited, underscoring the need for groundwater models for decision-making. Conventional numerical models solve physics-based partial differential equations at every node of a meshed geometry. Conversely, machine learning (ML) algorithms rely on mathematical equations between inputs and outputs by iteratively adjusting the model’s parameters. Each method has its strengths, but a key advantage of ML models is their ability to predict GWLs without requiring the calibration of uncertain boundary conditions, extensive hydrological parameters, and human-induced factors. In this study, Provincial Groundwater Monitoring Network (PGMN) wells, the largest publicly available dataset for GWL data in Southern Ontario, were used to detect long-term trends using the Mann-Kendall test. Results show that 92% of wells exhibit statistically significant trends, indicating the non-stationarity of the GWLs. This study also assessed the effectiveness of ensembles of Artificial Neural Networks (ANNs) – a type of ML model – and compared it to linear regression models to predict monthly GWL in PGMN wells throughout the Barrie-Oro Moraine, an important hydrological area in Southern Ontario. The models incorporated data preprocessing techniques and lagged variables such as temperature, total precipitation, hydrometric levels, and autoregressive GWL data. Results indicate that linear regression is inappropriate since its assumptions, such as linearity and homoscedasticity, are violated, with test R2 less than 0.2 across wells. Conversely, ensembles of ANNs exhibit better performance than linear regression, particularly in single watershed modelling scenarios, capturing the non-stationarity of GWL data, with median R2 values generally ranging between 0.5 and 0.9. In addition, single-well models generally outperformed combined and aggregated modelling configurations, and DMS-based preprocessing generally showed improved performance compared to DSM approaches. These results highlight the usefulness of data-driven models in GWL prediction studies. ML techniques may improve predictive accuracy and help water resource managers develop effective policies for groundwater management.