Unsupervised Detection of Groundwater Storage Anomalies in Ghana Using GRACE Satellite Data

This study investigates groundwater storage anomalies in Ghana using GRACE-derived data from 2004–2024 combined with statistical analysis and unsupervised machine learning. Groundwater anomalies were standardized using Z-scores, while an ensemble-based Isolation Forest framework was applied for anomaly detection. The results revealed substantial temporal variability, with persistent groundwater deficits during 2004–2009 followed by increasing positive anomalies after 2018. A total of 12 anomalous months were identified, comprising 5 deficit and 7 surplus events, with the strongest anomalies associated with groundwater deficits. Spatial analysis showed more frequent deficit anomalies in northern Ghana and stronger surplus occurrence in southern regions. Comparison with statistical thresholds further indicated that the machine learning framework captured additional subtle deviations beyond conventional threshold-based methods. Overall, the integration of GRACE observations with unsupervised anomaly detection provides a practical framework for groundwater monitoring in data-scarce environments.

The study area is Ghana, a West African country located between latitudes 4°44′ N and 11°11′ N and longitudes 1°12′ E and 3°11′ W, covering approximately 238,535 km2. Ghana’s climate is tropical with bimodal rainfall in the south and unimodal rainfall in the north, with annual precipitation ranging from about 800 mm in the coastal savannah to over 2,000 mm in the southwestern forest zone. The hydrogeology consists of three main formations: the Precambrian crystalline basement complex (about 54%), the Voltaian sedimentary basin (about 45%), and coastal sedimentary formations. Groundwater supplies approximately 60–70% of potable water in Ghana, particularly in rural areas.

Groundwater storage (GWS) anomalies were estimated from GRACE terrestrial water storage (TWS) observations after removing other hydrological components, including soil moisture, surface water, snow water equivalent, and canopy water, using the formulation: GWS = TWS − (SM + SW + SWE + CW). Groundwater storage anomalies (GWSA) were computed for each grid cell relative to the temporal mean, and anomalies were standardized using Z-scores with thresholds of ±1 and ±2 indicating moderate and extreme anomalies, respectively. For national-scale analysis, groundwater anomalies were spatially aggregated by computing the mean anomaly across all grid cells within Ghana at each time step.

Anomaly detection was performed using an ensemble isolation-based learning approach (Isolation Forest) that identifies rare observations through recursive random partitioning of the feature space. The method operates on the principle that anomalous observations are statistically sparse and therefore require fewer partitions to become isolated compared to normal observations. The anomaly score is estimated as s(x, n) = 2(−E(h(x))/c(n)), where E(h(x)) denotes the expected path length of observation x across the ensemble, n is the sample size, and c(n) represents a normalization factor. Values approaching 1 indicate strong anomalous behavior, while values closer to 0 correspond to normal observations. Model implementation used the scikit-learn framework with a contamination parameter of 0.05, corresponding to an expected anomaly proportion of 5%, and a fixed random state of 42 for reproducibility.

The temporal evolution of GRACE-derived groundwater storage anomalies in Ghana from 2004 to 2024 exhibits pronounced seasonal oscillations superimposed on substantial interannual variability. Negative anomalies dominate the early portion of the record, particularly between 2004 and 2009, where several observations fall below the moderate anomaly threshold and approach extreme negative conditions. This period is followed by a transition toward more variable groundwater conditions between approximately 2010 and 2017, during which anomalies fluctuate around the long-term mean. From 2018 onward, the groundwater anomaly series shows a gradual shift toward predominantly positive anomalies, with several observations exceeding the moderate threshold and approaching extreme positive conditions.

The detected anomalous months occur during both groundwater deficit and surplus phases, indicating that the ensemble-based detection framework captures deviations across different hydrological regimes. A total of 12 anomalous months were identified, comprising 5 groundwater deficit events and 7 groundwater surplus events. Despite the larger number of surplus anomalies, the most intense anomalous events were associated with strong groundwater deficit conditions observed during 2006 and 2007. The spatial distribution of the three highest-ranked anomalous events shows that the August 2006 and September 2006 events exhibit relatively widespread negative groundwater conditions across large portions of the country, whereas the April 2007 event displays greater spatial heterogeneity.

The spatial distribution of groundwater anomaly occurrence frequencies reveals that the overall anomaly occurrence pattern remains relatively spatially uniform because the anomaly proportion was constrained by the contamination parameter. However, separating anomaly occurrence into groundwater deficit and surplus categories reveals more distinct regional patterns. Groundwater deficit anomalies occur more frequently across northern and central Ghana, whereas surplus anomalies are more pronounced in southern and southwestern regions. These spatial contrasts are broadly consistent with the hydroclimatic and geological variability across the country, with northern Ghana characterized by longer dry periods and comparatively lower annual rainfall, while southern Ghana experiences wetter climatic conditions and more sustained recharge potential.

Groundwater storage variability is spatially heterogeneous across Ghana, with stronger temporal fluctuations observed in central and southwestern regions and comparatively lower variability across parts of northern Ghana. The confusion matrix comparing anomaly classifications from the standardized Z-score method and the ensemble-based anomaly detection framework indicates that both methods exhibit similar behavior in identifying normal groundwater conditions, while the machine learning approach additionally captures anomalous patterns that do not necessarily exceed fixed statistical thresholds, suggesting the presence of subtle or nonlinear groundwater variations embedded within the spatiotemporal anomaly structure.

The results demonstrate that groundwater storage dynamics in Ghana are strongly influenced by the combined effects of climatic variability, recharge processes, and regional hydrogeological conditions. Northern Ghana, characterized by longer dry periods and relatively lower recharge potential, exhibits more persistent deficit conditions and reduced temporal variability. In contrast, southern and southwestern regions display stronger groundwater fluctuations and more frequent surplus anomalies, potentially associated with wetter climatic conditions and enhanced recharge dynamics within weathered and fractured aquifer systems.

An important limitation is the coarse effective spatial resolution of GRACE observations (300–400 km). Given the relatively small size of Ghana and the heterogeneous distribution of its geological formations, aquifer systems, rainfall regimes, and groundwater-use pressures, groundwater variations occurring at local or sub-regional scales may be averaged within individual GRACE footprints. The spatial patterns should therefore not be interpreted as resolving individual aquifers or localized groundwater processes, but rather as broad-scale patterns in the GRACE-derived groundwater storage signal. The spatial interpolation applied to the GRACE-derived grid data improves visualization but does not add independent spatial information beyond the effective resolution of the satellite observations.

The study concludes that the ensemble-based anomaly detection framework successfully identified both extreme and subtle groundwater anomalies beyond conventional threshold-based analysis, demonstrating the potential of unsupervised learning techniques for groundwater monitoring in data-scarce environments. The comparative analysis further showed that machine learning and statistical approaches provide complementary perspectives for characterizing groundwater variability. Overall, the integration of GRACE observations with data-driven anomaly detection offers a practical framework for regional groundwater assessment and may support improved monitoring of hydrological extremes and groundwater resource management in Ghana and similar hydroclimatic regions.

Improvements to AI Systems:

1. Ensemble-based anomaly detection with contamination calibration: Implement an Isolation Forest framework with an adjustable contamination parameter (e.g., 0.05) and fixed random state for reproducibility. This allows AI systems to detect rare hydrological events without prior labeled data, capturing both extreme and subtle deviations that threshold-based methods miss.

2. Hybrid statistical-ML anomaly fusion: Combine Z-score standardization (thresholds ±1, ±2) with unsupervised ML outputs to create a dual-signal anomaly scoring system. The AI can classify anomalies as threshold-confirmed or ML-only (subtle/nonlinear), improving detection of early-warning signs of groundwater stress.

3. Spatiotemporal anomaly decomposition: Enhance AI systems to separate anomaly occurrence into deficit vs. surplus categories and map their spatial frequencies. This enables regional pattern recognition—e.g., identifying northern Ghana as deficit-prone and southern regions as surplus-prone—to support targeted water resource alerts.

4. Temporal trend segmentation: Use the observed 2004–2009 deficit and post-2018 surplus phases to train AI models for regime-shift detection. The system can automatically identify prolonged dry/wet periods and flag transitions, improving long-term groundwater forecasting.

5. Resolution-aware spatial interpolation: Integrate GRACE’s coarse resolution (300–400 km) as a known uncertainty constraint. The AI can apply interpolation for visualization but must explicitly avoid over-interpreting local aquifer details, instead providing confidence intervals for regional-scale assessments.

6. Complementary model comparison via confusion matrices: Build AI systems that automatically compare ML-based anomaly labels against statistical thresholds, quantifying agreement and divergence. This allows the system to learn which anomaly types are best captured by each method and adaptively weight them for future predictions.

What the Improved AI System Can Do:

• Real-time groundwater anomaly monitoring for data-scarce regions, flagging both extreme and subtle storage changes monthly.

• Regional early-warning alerts for droughts (northern Ghana) and floods/recharge surges (southern Ghana) based on spatial deficit/surplus patterns.

• Regime-shift detection to identify transitions from prolonged deficits to surpluses, aiding adaptive water management.

• Explainable anomaly reporting that distinguishes threshold-based events from ML-detected nonlinear deviations, providing actionable insights for hydrologists.

• Scalable transferability to other small, hydroclimatically diverse countries using GRACE or similar satellite data, with built-in resolution uncertainty handling.

Groundwater variability in Ghana remains poorly characterized due to limited long-term in-situ observations. This study investigates groundwater storage anomalies using GRACE-derived data from 2004-2024 combined with statistical analysis and unsupervised machine learning. Groundwater anomalies were standardized using Z-scores, while an ensemble-based Isolation Forest framework was applied for anomaly detection. The results revealed substantial temporal variability, with persistent groundwater deficits during 2004-2009 followed by increasing positive anomalies after 2018. A total of 12 anomalous months were identified, comprising 5 deficit and 7 surplus events, with the strongest anomalies associated with groundwater deficits. Spatial analysis showed more frequent deficit anomalies in northern Ghana and stronger surplus occurrence in southern regions. Comparison with statistical thresholds further indicated that the machine learning framework captured additional subtle deviations beyond conventional threshold-based methods. Overall, the integration of GRACE observations with unsupervised anomaly detection provides a practical framework for groundwater monitoring in data-scarce environments.