Assessing groundwater level modelling using a 1-D … – HESS

Understanding the Importance of Groundwater Level Forecasting

Groundwater is an invaluable natural resource, providing drinking water, irrigation, and industrial supplies across the globe. Monitoring and forecasting groundwater levels has become increasingly crucial in recent years as we face the challenges of climate change, growing water demands, and the need for sustainable groundwater management. Accurate groundwater level (GWL) predictions allow us to identify over-exploitation trends, assess water availability, and delineate potential soil subsidence zones – all essential for ensuring long-term water security.

Traditional physics-based and numerical groundwater models require extensive data on aquifer properties, geology, and topography, making them resource-intensive to develop and maintain. In contrast, data-driven machine learning (ML) approaches, such as artificial neural networks (ANNs), have emerged as powerful tools for GWL forecasting. These ML models can capture the non-linear dynamics of aquifer systems while requiring fewer input parameters. However, a key challenge remains in understanding the “black-box” nature of these models and the factors influencing their performance.

In this comprehensive article, we explore the application of a 1-D convolutional neural network (CNN) for GWL modelling across hundreds of observation wells in Lower Saxony, Germany. By linking the CNN model’s performance to geospatial characteristics and time series features of the monitoring sites, we aim to provide valuable insights into the complex interplay between groundwater dynamics and the factors driving them.

Leveraging the Power of CNNs for Groundwater Level Forecasting

Convolutional neural networks (CNNs) have emerged as a popular choice for GWL forecasting due to their flexibility, calculation speed, and reliable performance. Unlike traditional feedforward neural networks, CNNs are designed to effectively capture spatial and temporal patterns in data, making them well-suited for modelling the complex dynamics of groundwater systems.

In our study, we apply a 1-D CNN architecture to simulate monthly GWL time series at over 500 observation wells across Lower Saxony. The model uses precipitation (P) and temperature (T) as the primary input variables, with the network architecture and hyperparameters tuned individually for each well. This site-specific approach allows us to account for the unique characteristics of each groundwater system, rather than relying on a one-size-fits-all model.

The performance of the CNN models is evaluated using two widely adopted metrics: the coefficient of determination (R²) and the Nash-Sutcliffe efficiency (NSE). These metrics provide a comprehensive assessment of the models’ ability to accurately capture the observed GWL patterns.

Exploring the Influence of Geospatial and Time Series Features

While the CNN models demonstrate high performance at certain locations, we observed significant variations in accuracy when applying the same architecture across the regional study area. This discrepancy in model performance prompted us to investigate the potential factors driving these differences.

We focused our analysis on two key aspects: geospatial characteristics and time series features of the monitoring sites.

Geospatial Characteristics

The geospatial features examined in this study include:

• Land cover: The predominant land use type (e.g., non-irrigated arable land, forests, urban areas) surrounding the groundwater well can influence infiltration, recharge, and evapotranspiration processes, which in turn affect GWL dynamics.

• Proximity to waterworks: Wells located within the influence area of water extraction facilities may exhibit GWL patterns heavily influenced by managed abstraction, rather than responding primarily to climatic inputs.

• Leaf Area Index (LAI): This measure of vegetation density can indicate the extent of precipitation interception and evapotranspiration, potentially disrupting the direct relationship between rainfall and groundwater recharge.

• Topographic Wetness Index (TWI): This index reflects the landscape’s tendency to accumulate or transport water, which can impact the groundwater system’s sensitivity to meteorological drivers.

By correlating these geospatial features with the CNN model performance, we aimed to identify the physical factors that may enhance or hinder the models’ ability to accurately simulate GWL patterns.

Time Series Characteristics

In addition to the geospatial attributes, we also examined various time series features of the GWL observations, including:

• Autocorrelation: The degree of self-similarity within the GWL time series, which can indicate the aquifer’s responsiveness to climatic inputs.
• Flat spots: Periods of relatively constant GWL values, which may suggest an aquifer’s limited sensitivity to climate variability.
• Number of peaks: The complexity of the GWL signal, which could be linked to the strength of the precipitation-groundwater relationship.
• Approximate entropy: A measure of time series irregularity, reflecting the degree of unpredictability in the GWL dynamics.

These time series characteristics provide insights into the intrinsic properties of the groundwater systems, which may influence the CNN model’s ability to accurately capture the observed GWL patterns.

Key Findings and Insights

Our analysis of the CNN model performances revealed several important insights:

1. Proximity to Waterworks: Wells located in the vicinity of water extraction facilities exhibited lower model performances, likely due to the significant influence of managed groundwater abstraction on the observed GWL patterns. These anthropogenic impacts can disrupt the direct relationship between climatic inputs and groundwater levels, hindering the models’ ability to accurately simulate the observed dynamics.

2. Vegetation Density: Increased leaf area index (LAI), indicating denser vegetation cover, was associated with reduced model performance. This finding suggests that the interception and evapotranspiration processes driven by vegetation can weaken the link between precipitation and groundwater recharge, making it more challenging for the CNN models to capture the observed GWL patterns.

3. Groundwater Level Variability: GWL time series with more complex and irregular patterns, characterized by a higher number of peaks and greater approximate entropy, were generally better predicted by the CNN models. This observation indicates that the models perform better when the groundwater dynamics exhibit a stronger correlation with precipitation dynamics, allowing the models to learn these patterns more effectively.

4. Continuous Deviations from the Mean: Prolonged periods of GWL measurements above or below the mean negatively impacted model performance. These sustained deviations from the expected GWL patterns may be associated with managed aquifer recharge, excessive groundwater abstraction, or natural climate variability, all of which can disrupt the models’ ability to learn the relationships between climatic inputs and groundwater levels.

These insights highlight the importance of considering both geospatial and time series characteristics when developing and evaluating data-driven GWL models. By understanding the physical factors influencing groundwater systems, we can better interpret the performance of these “black-box” models and identify the limitations or strengths of using solely meteorological variables as inputs.

Implications and Future Directions

The findings from this study have several important implications for groundwater management and the application of data-driven modelling techniques:

1. Incorporating Explainable AI: While CNNs and other deep learning models have demonstrated impressive performance in GWL forecasting, their inherent “black-box” nature can hinder the understanding of the physical processes driving the observed groundwater dynamics. Integrating explainable AI (xAI) techniques, such as SHAP or LIME, can shed light on the non-linear relationships between model inputs and outputs, providing valuable insights for hydrogeologists and water resource managers.

2. Accounting for Anthropogenic Impacts: The significant influence of water extraction facilities on GWL patterns highlights the need to incorporate additional variables, such as pumping rates, into the modelling framework. By including these anthropogenic factors, data-driven models can better capture the complex interactions between climatic and human-induced drivers of groundwater dynamics.

3. Improving Model Architectures: While the 1-D CNN model used in this study demonstrated acceptable performance for a majority of the wells, the observed limitations in capturing certain GWL patterns suggest that more complex or tailored model architectures may be required. Exploring alternative deep learning structures, such as long short-term memory (LSTM) networks or hybrid models, could lead to further improvements in GWL forecasting accuracy.

4. Expanding Spatial and Temporal Scales: This study focused on a regional-scale analysis in Lower Saxony, Germany. Applying a similar framework to larger geographic areas or different climatic regions could provide valuable insights into the transferability and scalability of the observed relationships between model performance and geospatial/time series features.

As groundwater resources continue to face growing pressures, the need for reliable and interpretable forecasting models becomes increasingly critical. By bridging the gap between data-driven techniques and physical hydrogeological understanding, this study paves the way for more robust and informed groundwater management strategies, ultimately supporting the long-term sustainability of this vital resource.