Harnessing Remote Sensing and AI to Secure Groundwater in Morocco’s Dry Belt

The study by Ibn Tofail University researchers shows that machine learning, especially LightGBM, can accurately predict groundwater storage in Morocco’s semi-arid Rabat–Salé–Kénitra region using satellite and climate data. It highlights AI’s potential to guide sustainable water management under climate change pressures

A study carried out by researchers from Ibn Tofail University in Kenitra, Morocco, representing the Department of Computer Science, the Faculty of Humanities and Social Sciences, and the Laboratory of Plant, Animal, and Agro-Industry Productions confronts one of the most pressing challenges of the Rabat–Salé–Kénitra region: the growing depletion of groundwater in a semi-arid environment under the dual pressures of climate change and human exploitation. The region is an economic powerhouse of Morocco, supporting nearly 4.6 million people and contributing more than 15 percent of the country’s GDP, yet it faces mounting water stress. Over-extraction for agriculture, nitrate contamination, seawater intrusion in coastal aquifers, and untreated industrial discharges have combined to threaten both the quality and availability of groundwater. Conventional well-based monitoring of groundwater levels provides only a narrow view of this crisis. To overcome these limits, the researchers embraced a novel strategy, merging satellite-based remote sensing with advanced machine learning to forecast groundwater storage variability with daily resolution.

From Satellites to Algorithms: Building a Unique Dataset

The backbone of the study is a twelve-year dataset spanning 2010 to 2022, collected from nine sites that represent urban, agricultural, and coastal zones of the region. The team narrowed their analysis to three core variables most influential for aquifer behavior: groundwater storage anomalies derived from NASA’s GRACE satellites, precipitation records from the CHIRPS dataset, and average surface temperature data from MODIS instruments. These variables were meticulously preprocessed to eliminate missing values, correct outliers, and normalize scales, yielding a clean dataset of about 42,000 daily observations. With these inputs in hand, the researchers deployed six machine learning regression models: Decision Tree, Random Forest, Light Gradient Boosting Machine (LightGBM), CatBoost, Extreme Learning Machine (ELM), and Artificial Neural Networks (ANN). To ensure fairness in comparison, each model was tuned using GridSearch with cross-validation, a process that systematically identifies the best combination of hyperparameters to maximize predictive performance.

Winners and Losers in the AI Contest

The performance evaluation relied on three trusted indicators: the Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R²). The results revealed a sharp divide between strong and weak performers. LightGBM emerged as the standout, achieving an extraordinary R² of 0.9997 and an impressively low RMSE of 3.07. Its efficiency in handling large datasets and nonlinear interactions made it the most robust option. Random Forest and Decision Trees also proved reliable, with R² values of 0.9998 and 0.9996, respectively, though they required more computational resources. CatBoost demonstrated solid performance in managing categorical data but fell short of LightGBM’s precision. At the other end of the spectrum, Artificial Neural Networks and the Extreme Learning Machine struggled to generalize effectively, producing much higher error rates and requiring training times several times longer than the other models. Their underperformance, the authors argue, stems from their sensitivity to data normalization and their inability to fully capture the complex temporal patterns of groundwater fluctuations.

Forecasting Groundwater at the Local Scale

Beyond overall performance, the study examined how the models behaved at specific sites across the region, including Kenitra, Khemisset, Tiflet, and Al Kansera. Once again, LightGBM proved the most reliable, consistently tracking seasonal cycles of aquifer recharge and depletion and accurately mirroring sudden shifts following rainfall events. CatBoost and Random Forest displayed strong results too, though with slightly larger margins of error. Graphs included in the study highlight how LightGBM predictions almost perfectly align with real-world data, while ANN and ELM deviate significantly, sometimes predicting swings in groundwater levels that did not occur. This comparative analysis underscores how crucial model choice is when designing practical forecasting systems. Accuracy at daily resolution is especially important in semi-arid regions, where water scarcity can intensify in just a matter of weeks, and where miscalculations could mislead policymakers about the urgency of conservation measures.

A Roadmap for Policy and Future Research

The implications of the study reach far beyond technical exercises. Accurate and timely groundwater predictions can guide water managers in setting extraction limits, launching recharge initiatives, and balancing agricultural, industrial, and domestic demand. They also feed into broader environmental planning, as groundwater fluctuations are linked to geohazards such as landslides and soil instability. By embedding models like LightGBM into decision support systems, Moroccan authorities could create early warning frameworks and adaptive allocation strategies that respond to real-time changes. Importantly, because the approach relies on open-access satellite data, it is both cost-effective and scalable, offering a blueprint for other water-stressed regions worldwide. Still, the researchers acknowledge important limitations. The reliability of forecasts depends on data quality, which can be inconsistent across locations. The study also focuses on a narrow set of variables and models, leaving out socio-economic drivers, land use patterns, and deeper hydrological processes such as transmissivity and infiltration rates. Future research, they argue, should integrate deep learning models such as convolutional neural networks or transformer-based systems, while also combining statistical learning with physically based hydrological simulations. This hybrid approach would enhance interpretability, scientific rigor, and adaptability to future climate scenarios.

The work of Ibn Tofail University’s research team demonstrates how artificial intelligence can become a cornerstone of groundwater governance in semi-arid Morocco. By merging the global view of satellites with the predictive power of machine learning, they have produced models that allow policymakers to look ahead, anticipate shortages, and respond before aquifers are pushed beyond recovery. Among the tested algorithms, LightGBM stands out as the most accurate and efficient, setting a new benchmark for real-time water monitoring. In a region where economic vitality and social stability depend so heavily on groundwater, such advances may prove decisive in shaping a sustainable future.