Predicting groundwater level based on remote sensing and machine learning: a case study in the Rabat-Kénitra region

Groundwater is an essential resource for fulfilling water demands, sustaining industrial growth, supporting agricultural production, and preserving ecosystems on a global scale (Li et al. 2023). In arid and semi-arid regions, where surface water resources are scarce or unreliable, groundwater serves as an essential source of water for human survival and economic activities. Globally, groundwater accounts for nearly 30% of the world’s freshwater supply and supports over 2 billion people, making it indispensable for both current and future generations (Davamani et al. 2024). The significance of groundwater is further underscored by its role in mitigating the impacts of climate change on water availability, particularly in regions where climate-induced changes in precipitation patterns and increased evapotranspiration are expected to exacerbate water scarcity. Climate change poses significant challenges to groundwater recharge by altering precipitation patterns, increasing evapotranspiration, and intensifying droughts. These changes threaten the sustainability of groundwater resources and have profound consequences for critical and sensitive ecosystems (Loucks 2021). Additionally, groundwater over-extraction in response to declining surface water availability exacerbates the depletion of this vital resource, posing severe risks not only to ecological stability but also to human livelihoods and food security in affected regions.

These global challenges are particularly acute in Morocco, situated in North Africa, where water scarcity and the impacts of climate change are severe (Naghibi et al. 2015). Morocco faces inadequate water availability during critical periods, despite having underground reservoirs theoretically capable of meeting agricultural, industrial, and domestic needs (Alley 2009). However, these reservoirs are increasingly vulnerable to overexploitation and insufficient recharge, a situation exacerbated by the impacts of climate change (Swain et al. 2022). The consequences of water scarcity have become increasingly evident in Morocco, particularly in regions such as Rabat-Salé-Kénitra, where the pressures on groundwater resources continue to escalate. These challenges underscore the urgent need for advanced and sustainable groundwater management practices in Morocco, especially in light of the region’s growing population and economic demands.

Traditionally, detecting groundwater has involved costly and time-consuming fieldwork, such as borehole drilling and well construction, which provide limited spatial insights (Wang et al. 2018; Anomohanran et al. 2021). In response to these challenges, researchers are increasingly turning to advanced techniques, including artificial intelligence, machine learning models, and remote sensing, as promising alternatives to traditional methods (Tao et al. 2022). These approaches not only offer broader spatial coverage but also enable the integration of diverse datasets, thereby improving the accuracy and efficiency of groundwater predictions. By leveraging data from GRACE and MODIS satellites, researchers can derive environmental variables such as land surface temperature, soil moisture, terrestrial water storage, precipitation, and vegetation indices. The convergence of remote sensing and machine learning has revolutionized our understanding of groundwater resources, allowing for enhanced integration of ground observations and satellite data into comprehensive databases, which improves global insights into groundwater dynamics (Rodell et al. 2009; Rateb et al. 2020).

Groundwater is a critical resource for sustaining water supply, particularly in arid and semi-arid regions where surface water resources are limited or unreliable. In these regions, groundwater acts as a lifeline, supporting agricultural production, industrial operations, and domestic consumption, thereby playing a pivotal role in socioeconomic development. The sustainable management of groundwater is essential not only for meeting current water demands but also for ensuring the long-term resilience of ecosystems and communities. With the ongoing challenges posed by climate change, including more frequent droughts and shifts in precipitation patterns, the reliance on groundwater in arid regions is expected to increase, making its careful management a matter of urgent priority. Recent studies underscore the need for integrated water resource management approaches that prioritize groundwater sustainability to mitigate the risks associated with water scarcity and ecosystem degradation (Li et al. 2023; Noori et al. 2023; Sarkar et al. 2024; Yi et al. 2024). However, groundwater systems in these regions face increasing pressure due to climate change, which exacerbates water scarcity by significantly altering precipitation patterns, increasing the frequency and intensity of droughts, and elevating evapotranspiration rates. These changes disrupt the natural replenishment of aquifers, leading to decreased groundwater levels and heightened vulnerability of water resources. In regions already facing water scarcity, such as arid and semi-arid areas, the compounded effects of these climate-driven alterations can result in severe and prolonged water shortages, threatening both ecological stability and human water security (Noori et al. 2023; Mahdian et al. 2024). The consequences of these changes are particularly severe for ecosystems that rely on consistent groundwater recharge. Reduced recharge rates can lead to the desiccation of wetlands, the decline of groundwater-fed streams, and the loss of critical habitats that support a wide range of biodiversity. These impacts not only threaten the survival of numerous plant and animal species but also disrupt ecosystem services that are vital for human livelihoods, such as water purification, flood regulation, and climate stabilization. The degradation of these ecosystems can trigger a cascade of negative effects, further exacerbating the vulnerability of both natural and human systems to climate change (Mahdian et al. 2024).

Recent studies have highlighted significant declines in groundwater recharge across various regions, including Iran, where groundwater depletion has reached alarming levels. Noori et al. (2023) demonstrated that Iran’s groundwater recharge has decreased by approximately 3.8 mm/yr nationwide, primarily due to unsustainable water management practices compounded by climatic changes. Similarly, in Bangladesh, future groundwater potential mapping under various climate change scenarios has revealed substantial regional variability, emphasizing the importance of adaptive water management strategies (Sarkar et al. 2024). Yi et al. (2024) also investigated the impacts of climate variability on groundwater levels in South Korea, highlighting the need for advanced predictive models to mitigate the adverse effects of these changes.

Despite these advancements in understanding the challenges facing groundwater resources, significant gaps remain in the literature, particularly concerning the integration of machine learning models with diverse environmental datasets. Traditional methods, while valuable, often fall short in capturing the complex and non-linear relationships inherent in groundwater systems. For instance, Naghibi et al. (2015) employed basic regression models for groundwater level prediction, but these models exhibited limitations in handling the intricate interplay between multiple environmental variables, leading to reduced predictive accuracy.

Furthermore, Anomohanran et al. (2021) highlighted the constraints of traditional field methods, such as the labor-intensive nature of data collection and the challenges associated with large-scale or real-time monitoring. These methods, although accurate in localized studies, are not scalable for broader applications, particularly in regions with diverse hydrogeological conditions.

Recent advancements have explored the potential of machine learning techniques, offering promising alternatives to traditional methods. Tao et al. (2022) demonstrated the effectiveness of Support Vector SVR) and Random Forest (RF) models in predicting groundwater levels, leveraging remote sensing data to enhance model inputs. However, these studies frequently encounter challenges related to data integration, where inconsistencies between datasets can hinder model performance and generalization.

Moreover, while Rodell et al. (2009) and Rateb et al. (2020) underscored the critical importance of integrating diverse datasets such as satellite-based observations, climate data, and in-situ measurements to improve the robustness of predictions, a significant gap persists in the literature. Specifically, there is a lack of comprehensive approaches that combine multiple machine learning models with large-scale, high-resolution environmental data to improve prediction accuracy and reliability.

This study aims to address these gaps by integrating advanced machine learning models with extensive remote sensing and environmental datasets to develop a robust framework for predicting groundwater levels in the Rabat-Kenitra region.

This study aims to address the gaps in the current research by:

Integrating Diverse Environmental Variables: Utilizing data from GRACE and MODIS satellites to include a broad range of environmental variables such as land surface temperature, soil moisture, terrestrial water storage, precipitation, and vegetation indices.

Applying Advanced Machine Learning Models: Employing a range of machine learning models, including Gradient Boosting Regression (GBR), SVR, RF, and Decision Tree (DT) models, to handle complex and non-linear relationships between variables.

Focusing on a Region-Specific Case Study: Enhancing the prediction of groundwater levels specifically in the Rabat-Salé-Kénitra region, providing valuable insights and data-driven recommendations for sustainable groundwater management.

The novelty of this work lies in its comprehensive approach to integrating diverse environmental datasets with advanced machine learning techniques, which has not been extensively explored in the context of groundwater level prediction. By combining these methods, this research aims to improve the accuracy and reliability of groundwater predictions, contributing to more effective water resource management strategies.

The Rharb Aquifer, a crucial water source in the Rabat-Kénitra region, is characterized by its complex geology and hydrology. It comprises a sequence of alluvial deposits and permeable sandstone formations interbedded with less permeable clay layers (Rimi et al. 2006). The primary recharge processes involve infiltration from precipitation and lateral inflow from the surrounding Gharb Basin margins. The aquifer system is also influenced by the geomorphology of the region, including river networks that contribute to both recharge and discharge dynamics.

Historical data indicates that groundwater levels in the Rharb Aquifer have experienced significant fluctuations over the past decades. These fluctuations show a general decline in groundwater levels, attributed to increased extraction for agricultural and urban use (El Haouari & Khattabi 2012). The region has also seen rising groundwater demand driven by population growth and agricultural expansion, leading to over-extraction and subsequent concerns about sustainability.

The Rabat-Kénitra region faces several challenges related to groundwater management:

Over-Extraction: Intensive agricultural practices and urban development have led to excessive groundwater extraction, surpassing natural recharge rates. Contamination: Agricultural runoff and industrial pollutants pose risks to groundwater quality. Climate Change: Variability in precipitation patterns and increasing temperatures impact recharge rates and groundwater availability. Socioeconomic Impacts: Water scarcity affects local communities, leading to conflicts over resource allocation and necessitating robust management policies.

In this study, we employ a variety of machine learning models, including GBR, SVR, RF, and DT models. These models are trained using environmental variables derived from GRACE and MODIS satellite data. The reasons for choosing these methods include their robustness in handling complex and non-linear relationships, their ability to integrate diverse datasets, and their proven effectiveness in previous studies. The detailed methodology and performance evaluation of these models are discussed in the subsequent sections.

This section provides comprehensive information on the research area, data sources, modeling techniques, and assessment measures. The methodological approach employed in this research is time series regression using machine learning techniques (Aderemi et al. 2023). The performance of these models is evaluated using estimated coefficients such as Root Mean Squared Error (RMSE), R-squared, Mean Squared Error (MSE), and Mean Absolute Error (MAE).

The Rabat-Salé-Kénitra region, established in 2015 as one of the 12 regions of Morocco, includes the nation’s capital. This region resulted from the merger of the former regions of Gharb-Chrarda-Beni-Hssen and Rabat-Salé-Zemmour-Zaer. It is geographically bounded by Tanger-Tetouan-Al Hoceima to the north, Fès-Meknès to the east, Casablanca-Settat to the south, Beni-Mellal-Khénifra to the southeast, and the Atlantic Ocean to the west. The region spans a total area of 17,570 km², accounting for 2.5% of Morocco’s total area (hcp 2012).

The climate in Rabat-Salé-Kénitra is Mediterranean, characterized by chilly and wet winters, with average temperatures ranging from 0 to 5 degrees Celsius at night and reaching up to 17 degrees Celsius during the day. Summers are marked by an oceanic climate, with nights cooled by ocean moisture and daytime temperatures around 30 degrees Celsius. During spring and summer, the ‘Chergui’ desert winds can occasionally raise temperatures to 40 degrees Celsius for a few days (Hakimi & Brech 2021).

The region’s groundwater resources are sustained by several key aquifers, each with distinct geological and hydrological characteristics:

Maamora Aquifer: This extensive aquifer, primarily composed of Quaternary sands and limestones, serves as a crucial water reservoir with substantial storage capacity. It supports both agricultural and urban water needs, highlighting its importance for the region’s water supply (Zouhri 2001).

Bouregreg Coastal Aquifer: Located along the Atlantic coast, this aquifer is significantly influenced by marine conditions. It provides essential water resources to coastal cities such as Rabat and Salé, playing a vital role in supporting urban populations and industrial activities (Mahdaoui et al. 2024).

Gharb Aquifer: Situated in the Gharb plain, this aquifer consists of highly permeable alluvial deposits ideal for irrigation. It is extensively utilized for agricultural purposes, significantly contributing to the region’s agricultural productivity (Bouita Et Al 2021).

Tiflet Aquifer: Located inland from the coastal areas, the Tiflet Aquifer is characterized by sandstone and clay formations. It is critical for supplying water to rural communities, thereby enhancing the overall groundwater resources of the region (Faqihi et al. 2020).

Despite their critical importance, the aquifers in the Rabat-Salé-Kénitra region face several significant challenges. Overexploitation is a primary concern, as intensive agricultural and urban demands have led to groundwater extraction rates that exceed natural replenishment rates, resulting in depletion. Pollution also poses a substantial threat to groundwater quality, with contamination stemming from agricultural runoff, industrial effluents, and urban activities. Additionally, climate change exacerbates these issues by altering precipitation patterns and increasing temperatures, which in turn affect aquifer recharge rates and overall water availability.

To ensure the long-term sustainability of groundwater resources in the Rabat-Salé-Kénitra region, it is imperative to implement sustainable management practices. This includes robust monitoring and regulation of groundwater extraction to prevent overuse. Promoting water-efficient technologies in both agriculture and industry is essential to reduce water consumption. Furthermore, adopting integrated water resource management strategies that recognize the interconnectedness of surface water and groundwater systems is crucial for maintaining the balance and health of the region’s water resources.

The Rabat-Salé-Kénitra region is recognized for its significant groundwater reserves, with two primary aquifers of particular importance:

The Gharb Aquifer: Covering an area of 390 km², the Gharb aquifer boasts 126 million cubic meters (Mm³) of renewable water resources annually. This aquifer maintains a generally good water balance due to substantial recharge from precipitation and infiltration from the margins of the Gharb basin, highlighting its regional hydrogeological relevance (noa 2022).

The Maâmora Aquifer: Spanning approximately 4,000 km², the Maâmora aquifer is characterized as an unconfined aquifer, primarily recharged by rainfall infiltration. It serves as a substantial reservoir, providing 134 Mm³ of renewable water resources annually (noa 2022).

During the training validation phase, analysts assess how well different models have learned from their training sets by evaluating their performance on separate validation sets. This step allows for model adjustments and the exploration of alternative techniques. Through repeated testing and refinement, model improvements are achieved.

Once the best-performing model is identified, it undergoes evaluation with new data, referred to as the testing set. This case study utilizes four types of models: SVR, Random Forest, Decision Trees, and Gradient Boosting. Following rigorous testing, the final models are selected and deployed, culminating in their real-world application as depicted in the final part of the flowchart labeled Applied Model.

The daily data used in this study were collected from nine different sites in the Rabat-Salé-Kénitra region. The study utilized satellite data from the Moderate Resolution Imaging Spectroradiometer (MODIS) to evaluate a range of environmental factors in the region. Data were obtained from the NASA Earth Observing System Data and Information System (EOSDIS) data portal, covering a period from January 2010 to December 2022. The environmental factors analyzed include average surface temperature (AvgSurfT-tavg), groundwater storage (GWS-tavg), soil moisture (SoilMoist-S-tavg), terrestrial water storage (TWS-tavg), elevation, soil type, land surface temperature (LST), evapotranspiration (evap-Tr), normalized difference vegetation index (NDVI), and precipitation. The analysis also utilized GRACE data.

Groundwater storage represents the total volume of water stored underground in aquifers, averaged over a specific period. Data sourced from GRACE-DA, available from February 24, 2000, to February 17, 2023. Declines in groundwater storage reflect reductions in available groundwater, often due to excessive extraction for agricultural, industrial, and domestic purposes. Sustained over-extraction can lead to significant drops in groundwater levels, affecting water availability and quality (Rodell et al. 2009).

Precipitation is any form of water, liquid or solid, that falls from the atmosphere and reaches the ground. Data sourced from CHIRPS, available from February 24, 2000, to May 5, 2023. Precipitation is the primary source of groundwater recharge. Reduced precipitation can lead to lower groundwater levels, especially in regions that rely heavily on rainfall to replenish aquifers (Taylor et al. 2013).

Average surface temperature refers to the mean temperature recorded at the earth’s surface over a specified period. Data sourced from GRACE-DA, available from February 24, 2000, to February 17, 2023. Increased surface temperatures can lead to higher rates of evaporation, reducing the amount of water available to infiltrate into the ground and recharge aquifers. This reduction in infiltration can significantly exacerbate groundwater depletion, especially in arid and semi-arid regions (Kundzewicz et al. 2007).

Soil moisture refers to the amount of water present in the soil, averaged over a specific period. Data sourced from GRACE-DA, available from February 24, 2000, to May 5, 2023. Low soil moisture can reduce the amount of water that percolates down to recharge aquifers. During periods of drought, reduced soil moisture can lead to increased reliance on groundwater for irrigation, further depleting groundwater resources (Döll 2009).

Land surface temperature is the temperature of the land surface as measured by remote sensing instruments. Data sourced from MODIS, available from February 24, 2000, to May 5, 2023. High land surface temperatures can increase evapotranspiration rates, reducing soil moisture and groundwater recharge. This can lead to a more rapid depletion of groundwater reserves, especially in hot climates (Wang & Dickinson 2012).

Evapotranspiration is the sum of evaporation from the land surface and transpiration from plants. Data sourced from GRACE-DA, available from February 24, 2000, to May 5, 2023. High evapotranspiration rates can significantly reduce the amount of water available for groundwater recharge. In agricultural areas, high evapotranspiration can lead to increased irrigation demands, further stressing groundwater resources (Wada et al. 2010).

To provide a comprehensive overview of the dataset characteristics, we performed a statistical analysis of the key environmental parameters. Table 1 presents the mean, standard deviation (Std), minimum (Min), median (50%), and maximum (Max) values for each parameter.

This statistical analysis provides a clear understanding of the data distribution and variability for each environmental factor, which is crucial for interpreting the modeling results and understanding the influence of these factors on groundwater levels.

The monitoring dataset for this study was obtained from various sources, including NASA’s GRACE and MODIS satellites, and Google Earth. The dataset includes environmental variables such as land surface temperature, soil moisture, terrestrial water storage, precipitation, and vegetation indices. These data points were preprocessed through extraction, cleaning, and normalization before being used as predictors in the machine learning models. The use of Google Earth provided additional geospatial data to complement the satellite-derived variables.

Table 2 presents a comprehensive overview of environmental data sources and their availability for the Rabat-Salé-Kénitra region in Morocco. It includes various environmental factors such as average surface temperature, groundwater storage, soil moisture, terrestrial water storage, and others, along with their data sources, specific bands, and the periods for which the data is available.

Despite their critical importance, the aquifers in the Rabat-Salé-Kénitra region face several significant challenges. Overexploitation is a primary concern, as intensive agricultural and urban demands have led to groundwater extraction rates that exceed natural replenishment rates, resulting in depletion. Pollution also poses a substantial threat to groundwater quality, with contamination stemming from agricultural runoff, industrial effluents, and urban activities. Additionally, climate change exacerbates these issues by altering precipitation patterns and increasing temperatures, which in turn affect aquifer recharge rates and overall water availability.

The calibration of the machine learning models was performed using a combination of various metrics and hyperparameter tuning. Each model – SVR, DT, RF, and GBR – was trained and evaluated using metrics such as R-squared (R²), MAE, MSE, and RMSE. The process involved:

1. Model Training and Evaluation: The models were trained on the training dataset and evaluated on the test dataset. The performance of each model was measured using the aforementioned metrics.

2. Hyperparameter Tuning: The models underwent hyperparameter tuning to optimize their performance. For instance, the Gradient Boosting Regression model was fine-tuned by running a grid search with cross-validation, optimizing hyperparameters such as learning rate, number of estimators, and maximum tree depth.

One critical factor in determining the accuracy of predictions when using machine learning models is the proper preparation of input data. The preprocessing and normalization of variables can significantly impact the quality of insights gained, as some features may inherently hold more weight than others (Yu et al. 2018; Guzman et al. 2019; Mohammadi 2019). In this study, we examined 42,000 readings collected from multiple sites, all of which were thoroughly checked and cleaned prior to analysis. This dataset included up-to-date measurements of water levels and various other environmental factors, recorded daily.

To handle and analyze this extensive dataset, we utilized Python 3.10.11 and several essential libraries, including NumPy, SciPy, pandas, and Matplotlib. For machine learning tasks, such as SVR, DTs, and RFs, we employed the scikit-learn library. This library facilitated not only model fitting but also efficient cross-validation, a process that can be computationally expensive due to the repeated splitting of the dataset.

To effectively train and test our models without overfitting to the training data, we divided the dataset into two subsets: 70% for training and 30% for testing and validation (Davis et al. 1999). The concept of ‘generalization’ is crucial in this context; a well-trained model should perform equally well on both the training and testing datasets. Performance evaluation involves training the model on the training set (70% of the data) and then assessing its learning on the remaining 30%. To further fine-tune the models and avoid overfitting, researchers often employ a validation set drawn from the testing data. This additional step ensures that the model does not merely memorize the training data but can generalize its learning to new, unseen data.

In this study, we analyzed various environmental and meteorological factors that could potentially influence groundwater levels. The factors considered were: average surface temperature, soil moisture, terrestrial water storage, altitude, soil type, land surface temperature, evapotranspiration (the combined process of evaporation and plant transpiration), normalized difference vegetation index (NDVI) as an indicator of plant health, and precipitation. For each of these factors, we calculated the correlation coefficient with groundwater storage. Factors with low correlation values were excluded from further analysis. This approach enabled the identification of the most relevant features for estimating groundwater levels. Utilizing correlation analysis in this manner enhances the transparency and reliability of the study by providing well-founded justifications for the inclusion or exclusion of specific factors (Smith & Vasishth 2020).

Based on Figure 3, the following key observations can be made: Total terrestrial water storage exhibits the strongest positive correlations with groundwater levels, as indicated by the dark blue bars. Elevation and soil type also display moderately positive relationships with groundwater. In contrast, variables such as precipitation, evapotranspiration, and vegetation indices demonstrate negative correlations with groundwater levels. Additionally, land surface temperature and elevation show relatively strong negative correlations with groundwater levels.

This correlation analysis provides valuable insights into the complex interrelationships between various environmental factors and groundwater dynamics, which can be instrumental for groundwater management and decision-making processes.

Supplementary material, Figure A.11 illustrates the correlation between average surface temperature and groundwater levels, highlighting the potential influence of temperature on groundwater dynamics.

Supplementary material, Figure A.12 illustrates the relationship between total water storage (TWS) derived from GRACE satellite data and groundwater levels, suggesting that variations in terrestrial water storage can serve as a critical predictor for fluctuations in groundwater levels.

Supplementary material, Figure A.13 demonstrates the correlation between soil moisture, surface air temperature, and groundwater levels, emphasizing the potential influence of soil moisture and temperature conditions on groundwater recharge and depletion processes.

The correlations between groundwater levels and various factors are depicted in Supplementary material, Figures A.11, A.12, and A.13. These figures indicate that average surface temperature may impact groundwater dynamics, changes in terrestrial water storage (from GRACE satellite data) are significant predictors for groundwater level fluctuations, and soil moisture levels and temperatures could influence the recharge and depletion of aquifers. Figure 3 presents a matrix illustrating all these relationships, aiding in the identification of key factors to consider when developing our AI models.

In the methodology section, it was mentioned that correlation analysis was performed to identify the most relevant predictors. The specific statistical methods or techniques used for this analysis included: The Pearson Correlation Coefficient measures the linear correlation between two variables, providing a value between -1 (perfect negative correlation) and 1 (perfect positive correlation). This coefficient was used to determine the strength and direction of the relationships between each environmental factor and groundwater storage. A bar visualization was generated to illustrate the correlation matrix, allowing for quick identification of strong and weak correlations through color coding. These techniques ensured that the selection of predictors was based on statistically significant relationships, thereby enhancing the robustness and reliability of the groundwater level predictions.

n this study, four machine learning models were employed to address groundwater-related tasks, including modeling, prediction, and forecasting. The selected models, DT, RF, Gradient Boosting Regression (GBR), and SVR, were trained on a prepared training set. Each model’s performance was subsequently evaluated using a test set, with metrics such as MSE, MAE, and R-squared (R²) serving as the primary indicators of accuracy. Among these models, Gradient Boosting Regression demonstrated superior performance. This model was further optimized through hyperparameter tuning using a grid search with cross-validation. The final evaluation, conducted on an independent dataset, confirmed the robustness of the optimized GBR model, which consistently outperformed the other models in accuracy.

SVR is a sophisticated machine learning technique designed for regression tasks, extending the principles of Support Vector Machines (SVM) to predict continuous values (Figure 4). The objective of SVR is to identify a function that maintains a deviation from the actual observed targets y within a margin of ε, also preserving the flatness of the function (Cortes & Vapnik 1995; Aghelpour et al. 2019).

A significant advantage of SVR is its flexibility in handling both linear and non-linear relationships between input variables and the target variable. This flexibility is achieved through the application of kernel functions, such as the Radial Basis Function (RBF), which enable the mapping of input features into high-dimensional space, thereby enhancing SVR’s performance with complex datasets (Chengcheng et al. 2021).

SVR is recognized for its robust performance in handling non-linear and high-dimensional data, making it a strong candidate for complex regression tasks (Kawashima & Kumano 2017; Lan & Xiao 2023; Habib et al. 2024). Nonetheless, the method’s computational intensity, coupled with its sensitivity to hyperparameter tuning and outliers, are significant limitations that need careful consideration.

GBR is a sophisticated ensemble learning technique that excels in predictive modeling by aggregating the contributions of multiple weak learners, typically decision trees. Unlike traditional regression approaches, GBR operates by sequentially fitting trees to the residuals or errors of previous models (Figure 5), effectively enhancing the model’s accuracy at each iteration. This method allows GBR to manage complex, non-linear relationships within the data, making it particularly useful for tasks where traditional linear models fall short (Chen & Guestrin 2016).

RF is a highly robust ensemble learning method that leverages the power of multiple decision trees to enhance predictive accuracy and model stability (Figure 5). Unlike single decision tree models, which can be prone to overfitting, RF mitigates this risk by averaging the predictions of numerous trees, each trained on a random subset of the data (Figure 6). This approach not only reduces variance but also enhances the model’s ability to generalize to new data, making it particularly well-suited for complex and high-dimensional datasets (Breiman 2001).

DTs are one of the most widely used and intuitive machine learning techniques, applicable to both classification and regression tasks. They are particularly valued for their simplicity and the clarity with which they visualize the decision-making process. This makes DTs an excellent choice for tasks where interpretability and transparency are crucial, such as in policy-making or educational contexts. Moreover, DTs are well-suited for handling complex, non-linear relationships between input features and the target variable, making them versatile tools for a wide range of applications (Safavian & Landgrebe 1991).

To provide a comprehensive understanding of the strengths and limitations of each machine learning model employed in this study, the following Table 3 presents a detailed summary of the key advantages and disadvantages associated with the models utilized, supported by recent studies:

The selection of an appropriate machine learning model for groundwater prediction involves balancing several factors, including predictive accuracy, computational efficiency, and model interpretability. The table above highlights the strengths and weaknesses of each model, offering insights into their suitability for different aspects of groundwater modeling.

GBR is noted for its high predictive accuracy, which it achieves by iteratively minimizing residual errors. This makes GBR particularly effective for capturing complex, non-linear relationships in environmental data, which are common in groundwater systems. However, this accuracy comes at the cost of increased computational demands and a higher risk of overfitting if the model’s hyperparameters are not carefully tuned. The complexity of GBR models can also make them less interpretable, posing challenges in scenarios where model transparency is required.

RF offers a robust alternative, particularly in its ability to prevent overfitting through the ensemble method of averaging multiple decision trees. RF is well-suited for handling large datasets and provides valuable feature importance metrics, aiding in model interpretation. Nevertheless, the computational cost associated with constructing numerous trees can be significant, and the model’s complexity may reduce its interpretability compared to simpler models.

SVR is highly flexible, capable of modeling both linear and non-linear relationships with high-dimensional data. Its strength lies in its generalization ability, which enhances predictive reliability. However, SVR is computationally intensive and sensitive to the choice of hyperparameters, which can complicate its application, particularly in large-scale datasets.

DTs are perhaps the simplest and most interpretable model, making them an attractive option for initial analyses or when transparency is a priority. DTs are computationally efficient and capable of capturing non-linear relationships. However, they are prone to overfitting, especially with deep trees, and can be unstable, with small changes in the data potentially leading to significantly different tree structures. Additionally, DTs may exhibit bias towards features with more levels, affecting the accuracy of the model’s predictions.

In conclusion, the choice of model should be guided by the specific needs of the groundwater prediction task at hand. GBR is recommended when high accuracy and the ability to model complex relationships are prioritized, despite the computational costs. RF is ideal for large datasets and scenarios where overfitting must be minimized, although it may require substantial computational resources. SVR is best suited for high-dimensional data with a focus on generalization, though it demands careful hyperparameter tuning. Finally, DTs are most appropriate when interpretability and computational efficiency are paramount, with the caveat that their susceptibility to overfitting and instability must be managed. By understanding these tradeoffs, researchers can select the most appropriate model, or combination of models, to optimize groundwater predictions.

Table 4 compares different models of Gradient Boosting Regression. Each model has been fitted with a different set of hyperparameters: for example, learning rate, number of estimators, maximum tree depth, or random state. The table shows how the models perform according to four metrics–R-squared value, MAE, MSE, and RMSE.

With an R-squared (R²) score of 0.99, the Gradient Boosting Regression model performed well, explaining 99% of the variation in the target variable. The model had an MAE of 1.94, which means that the predicted values were 1.94 units off from the actual values on average. This model’s MSE was 8.87, indicating very minor squared discrepancies between predicted and actual values. The RMSE was 2.98, demonstrating a low average magnitude of prediction errors on the target variable’s original scale.

In Table 5, we assess how well different decision tree models perform under various hyperparameter configurations. These configuration settings cover maximum depth, minimum samples per leaf, and the number of features considered when splitting. Our evaluation criteria consist of four statistics for each model: R-squared, MAE, MSE, and RMSE. While R-squared indicates the proportion of variance in the dependent variable explained by the model, MAE, MSE, and RMSE all speak to how accurately these models make predictions; lower values for RMSE indicate better performance (as do higher ones for R-squared).

With an R2 score of 0.98, the Decision Tree model explained 98% of the variation in the target variable. When compared to the Random Forest and Gradient Boosting Regression models, the MAE of 3.05 suggests a greater average deviation. The Decision Tree model’s RMSE score was 4.93, suggesting a somewhat greater average squared difference between predicted and actual values.

In Table 6, we have compared how random forest models perform under different hyperparameters. These include maximum tree depth, minimum samples found at leaf nodes, and the most features to consider when splitting. We then assess the models using various criteria: R-squared value plus MAE, MSE or RMSE. Such measures help understand both how well your chosen factors explain variation in the dependent variable – as indicated by higher R-squared scores – along with their overall accuracy; lower values of MAE, MSE, and RMSE suggest better-performing models.

The Random Forest model produced a high R2 value of 0.97, matching the Decision Tree model. The MAE of 4.05 indicates a reasonably minimal average deviation from the actual values. The MSE of 31.67 suggests moderate squared deviations between anticipated and actual values.

In Table 7, we assess how three SVM models perform under different hyperparameters. These include the type of kernel, gamma value, and regularization parameter (C). Our evaluation uses various metrics: R² score, MAE, MSE, and its square root, RMSE.

The R² value of 0.95 produced by the SVR model indicates that it can account for 95% of the variation in the target variable. It did, however, have a higher MAE of 6.73, suggesting a greater average variance between anticipated and actual values. SVR had an MSE of 69.28, showing higher squared differences than the Gradient Boosting Regression model.

The performance of four different AI models (Gradient Boosting Regression, SVR, Random Forest, and Decision Tree) was evaluated for groundwater level prediction using various metrics, including R-squared R², MAE, MSE, and RMSE.

Table 8 displays that among all models used, Gradient Boosting for Regression performed the best based on three metrics: An R² of 0.99 indicates an extremely good fit – in other words, there is little difference between predicted values and actual ones; when it came to making mistakes (with its forecasts), this algorithm also had the smallest Mean Squared Error (8.87) as well as Mean Absolute Errors (1.94).

The figure presents a comparison of the R-squared, MAE, and RMSE values for the GBR, SVR, RF, and DT models.

The researchers performed extensive hyperparameter tuning for each of the models to optimize their performance. For the GBR model, they evaluated four different configurations with varying learning rates, number of estimators, and maximum depths. The best-performing GBR model had a learning rate of 0.1, 100 estimators, a maximum depth of 6, and a random state of 42. This model configuration achieved the highest R-squared of 0.99, the lowest MAE of 1.94, and the lowest RMSE of 2.98 compared to the other GBR settings.

The GBR model clearly outperformed the other models, with the highest R-squared value of 0.99, the lowest MAE of 1.94, and the lowest RMSE of 2.98. This indicates that the GBR model had the best fit between the predicted and observed groundwater levels, and the lowest average magnitude of errors in its predictions.

The DT model demonstrated the second-best performance, with an R-squared of 0.98, followed by the RF model with an R-squared of 0.97. Both the DT and RF models outperformed the SVR model, which had a lower R-squared of 0.95 and higher error values (MAE = 6.73, RMSE = 8.32).

The comprehensive hyperparameter tuning process and the visual comparison provided in the figure further reinforce the superior performance of the GBR model compared to the other machine learning techniques evaluated in this study. The clear separation in the performance metrics underscores the GBR model’s ability to accurately predict groundwater levels.

We performed cross-validation on the training set to assess the model’s reliability. The results of the cross-validation for each model are summarized in Table 9. The cross-validation results demonstrate that the GBR model consistently outperforms the other models across all evaluation metrics. The high R-squared value and low error metrics indicate the robustness and reliability of the GBR model in predicting groundwater levels. The low standard deviations in the cross-validation scores further emphasize the model’s stability and generalization capability.

The SVR model, while performing well, shows higher variability in its predictions, as indicated by the higher standard deviations in the cross-validation metrics. This suggests that the SVR model might be more sensitive to the specific training data subsets.

The RF and DT models also perform well, with the DT model showing slightly better performance in terms of MAE and RMSE compared to the RF model. However, both models exhibit higher error metrics compared to the GBR model, indicating that GBR is more suitable for this specific application.

Overall, the cross-validation results confirm the superior performance and reliability of the GBR model for groundwater level prediction, making it the preferred choice among the evaluated models.

The results demonstrate the effectiveness of AI models, particularly the GBR model, in predicting groundwater levels. High R² values and low error metrics suggest that these models can accurately capture the complex relationships between groundwater levels and factors such as weather patterns, remote sensing data, and soil moisture content.

The correlation numbers provide insights into the relationships between groundwater levels and various factors. For instance, a strong relationship with terrestrial water storage (TWS) suggests that incorporating more satellite data, such as GRACE, could enhance predictions. Similar correlations with soil moisture and surface air temperatures highlight their significant roles alongside weather information in modeling groundwater levels.

However, it is important to acknowledge the limitations of this research. A key constraint is the exclusion of all geological factors influencing groundwater levels in the study region. Geology plays a crucial role in groundwater dynamics, and its omission may have impacted the model’s accuracy. Future studies should incorporate geological information to provide a more comprehensive understanding of groundwater behavior.

Despite these limitations, the results of this study are valuable for land management. Accurate knowledge of groundwater levels can inform decisions on crop selection, irrigation planning, and resource allocation, especially in regions with limited water supplies. Understanding these levels is essential for sustainable land use practices.

Furthermore, this research underscores the importance of precise and comprehensive environmental data. Reliable measurements of soil moisture, temperature, precipitation, and other factors can improve groundwater level predictions and assist in areas with insufficient data. This information is crucial for understanding the impact of human activities and natural events on groundwater supplies and their interactions with surface water systems.

Groundwater level predictions can inform the development of comprehensive water management strategies at the regional or watershed level. Accurate forecasts of groundwater availability can help water authorities make informed decisions about water allocation, groundwater extraction limits, and infrastructure planning to ensure sustainable water supplies. For example, water managers could use the model outputs to determine safe yield limits for groundwater pumping and implement policies to restrict extraction in areas prone to depletion.

Farmers and agricultural agencies can utilize the groundwater level predictions to optimize irrigation scheduling and water usage. By knowing the expected groundwater levels, they can adjust irrigation practices to match crop water requirements, reducing over-extraction and improving the efficiency of water use. For instance, farmers could use the model outputs to determine the ideal timing and volumes of groundwater pumping for irrigation, ensuring crops receive the necessary water without depleting the aquifer.

The model outputs can help identify areas at risk of groundwater depletion, allowing farmers and policymakers to implement sustainable agricultural practices, such as crop selection, soil management, and precision irrigation techniques, to maintain long-term agricultural productivity. This information can guide the transition to water-efficient crops and the adoption of precision irrigation systems in vulnerable regions.

Groundwater level information can be incorporated into precision farming applications, enabling farmers to make data-driven decisions about irrigation, fertilizer application, and other agricultural inputs to maximize yields while minimizing environmental impacts. Farmers could integrate the model predictions into their digital farming platforms to optimize water use and improve the overall sustainability of their operations.

The model outputs can be used to assess the potential impacts of climate change, land use changes, or other environmental factors on groundwater resources, supporting the development of adaptive management strategies and environmental impact assessments. Researchers and environmental agencies can leverage the model predictions to evaluate the long-term sustainability of groundwater-dependent ecosystems and plan for mitigation or adaptation measures.

The results of this study, particularly the performance of the GBR model in predicting groundwater levels, were evaluated against existing studies employing various machine learning techniques. This comparative analysis provides insights into the relative strengths and weaknesses of different modeling approaches in similar contexts.

Study 1: (Feng et al. 2024) explored the application of traditional and deep machine learning algorithms, including Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), SVM, DT, RF, and Generative Adversarial Network (GAN) for groundwater level prediction in Izeh City, Iran. Their findings showed that the CNN model achieved the highest accuracy, with an RMSE of 0.0558 and an R² of 0.9948. The high performance of CNN is attributed to its robustness against noise and its capability to handle large datasets. In comparison, the GBR model in our study achieved an R² value of 0.99 and an RMSE of 2.98, indicating that traditional ensemble methods can still provide highly accurate predictions comparable to advanced deep learning models under certain conditions (Feng et al. 2024).

Study 2: (Ibrahem Ahmed Osman et al. 2021) conducted a study on groundwater level prediction in Selangor, Malaysia, using XGBoost, Artificial Neural Network (ANN), and SVM models. Their results indicated that the XGBoost model outperformed the others, with an RMSE of 0.138 and an R² of 0.920. Although XGBoost demonstrated promising results, the GBR model in our study exhibited a higher R² value, suggesting that ensemble methods like Gradient Boosting can sometimes offer superior performance in capturing complex relationships in groundwater data (Ibrahem Ahmed Osman et al. 2021).

Study 3: The study’s approach to groundwater level prediction using GBR aligns with existing literature, offering unique contributions relevant to regions with limited time series data. (Li et al. 2023) demonstrated GBR’s superiority in the Bilate watershed in Ethiopia, achieving an R-squared value of 0.77 with non-time series data, similar to our findings where GBR excelled due to its ability to handle non-linear relationships. This study also aligns with (Zhang et al. 2023), emphasizing the importance of variables like LST, NDVI, and precipitation.

Study 4: (Zhang et al. 2023) developed a real-time groundwater level forecasting strategy using hybrid models combined with remote sensing data, focusing on arid and semi-arid regions. They utilized Deep Learning algorithms and Ensemble Machine Learning models such as LSTM, RF, and XGB, enhanced with Wavelet Transformation (WT). Their findings showed that hybrid models outperformed standalone models, with WT-LSTM achieving the best performance. Similarly, our study employed GBR, SVR, RF, and DT models, with GBR showing the highest accuracy. The integration of diverse environmental variables like terrestrial water storage and soil moisture in our study provides a more comprehensive view of factors influencing groundwater levels, aligning with Zhang et al.’s emphasis on hybrid models. Additionally, Zhang et al. used the SHapley Additive exPlanations (SHAP) method for model interpretability, which parallels our correlation analysis and variable importance assessments. In conclusion, the findings of this study align with existing literature, emphasizing the effectiveness of machine learning models in groundwater level prediction. The superior performance of the GBR model, with an R² value of 0.99, underscores the potential of traditional ensemble methods in achieving high predictive accuracy. Additionally, this comparative analysis demonstrates that both traditional and deep learning models have unique strengths, and the choice of model can depend on specific dataset characteristics and prediction objectives.

In conclusion, the findings of this study align with existing literature, emphasizing the effectiveness of machine learning models in groundwater level prediction. The superior performance of the GBR model, with an R² value of 0.99, underscores the potential of traditional ensemble methods in achieving high predictive accuracy. Additionally, this comparative analysis demonstrates that both traditional and deep learning models have unique strengths, and the choice of model can depend on specific dataset characteristics and prediction objectives.

This study presents a comprehensive analysis of machine learning models for predicting groundwater levels, with a specific focus on the GBR model. The findings highlight the potential of these models in capturing the complex relationships between groundwater levels and various environmental factors.

The novelty of this work lies in its specific application and combination of methodologies:

Integration of GRACE and MODIS Data: This study uniquely integrates environmental variables from both GRACE and MODIS satellites to predict groundwater levels, which is not commonly done in existing studies.

Comprehensive Model Comparison: It provides a thorough comparison of multiple machine learning models (SVR, DT, RF, GBR) for groundwater level prediction, offering insights into their relative performances.

Focus on Rabat-Salé-Kénitra Region: The research addresses the specific challenges and conditions of the Rabat-Salé-Kénitra region in Morocco, contributing to localized water resource management strategies.

Environmental and Climatic Variable Analysis: The study includes an in-depth analysis of how various environmental and climatic variables influence groundwater levels, which can help in understanding the dynamics of groundwater systems in similar regions.

The key findings of this study underscore the effectiveness of the GBR model in predicting groundwater levels, with an R² value of 0.99, highlighting its superior performance compared to other models evaluated. This aligns with and expands upon existing literature, demonstrating that traditional ensemble methods can still provide highly accurate predictions comparable to advanced deep learning models under certain conditions.

The comparative analysis with existing studies reveals that while deep learning models such as CNN and EDL have shown high accuracy in some contexts, the GBR model’s robust performance, particularly in handling non-linear relationships and data noise, makes it a valuable tool for groundwater prediction in various settings.

Despite the promising results, this study has several limitations:

Exclusion of Geological Factors: The study does not consider geological factors influencing groundwater levels in the region, which could significantly impact model accuracy.

Data Availability and Quality: The accuracy of predictions is highly dependent on the quality and availability of input data. Limited or low-quality data could affect the model’s performance.

Generalization to Other Regions: While the models perform well in the Rabat-Salé-Kénitra region, their generalizability to other regions with different hydrogeological conditions needs further investigation.

Incorporating Geological Factors: Including geological information in the models to provide a more comprehensive understanding of groundwater behavior.

Advanced Machine Learning Techniques: Exploring advanced machine learning techniques such as deep learning and artificial neural networks to improve prediction accuracy.

Long-term Data Collection: Enhancing long-term data collection efforts to improve the quality and reliability of input data.

Impact of Land Use Changes: Investigating the effects of land use changes and ecosystem health on groundwater levels to build more robust models.

The results and methods of this study can be implemented in similar environmental problems as follows:

Water Resource Management: Groundwater level predictions can inform the development of comprehensive water management strategies at the regional or watershed level, helping water authorities make informed decisions about water allocation, groundwater extraction limits, and infrastructure planning. Sustainable Agriculture: Farmers and agricultural agencies can utilize groundwater level predictions to optimize irrigation scheduling and water usage, reducing over-extraction and improving water use efficiency. Environmental Impact Assessments: The model outputs can be used to assess the potential impacts of climate change, land use changes, or other environmental factors on groundwater resources, supporting the development of adaptive management strategies and environmental impact assessments.

In summary, this study provides valuable insights into the application of machine learning models for groundwater level prediction. The findings underscore the importance of precise and comprehensive environmental data and highlight the potential of these models in addressing water resource management and environmental sustainability challenges.

The authors declare there is no conflict.

All relevant data are included in the paper or its Supplementary Information.