Adam et al. (2026) Application of random forest modeling to evaluate groundwater storage changes in the Breede Water Management Area, South Africa

Identification

• Journal: Hydrogeology Journal
• Year: 2026
• Date: 2026-04-10
• Authors: Maal A. Adam, Stephanie Scheiber-Enslin, K. A. Ali
• DOI: 10.1007/s10040-026-03047-w

Research Groups

• School of Geosciences, University of Witwatersrand, Johannesburg, South Africa
• Department of Geology and Environmental Geoscience, College of Charleston, Charleston, SC, USA

Short Summary

This study integrated GRACE satellite observations, in situ groundwater levels, and GLDAS-derived hydrological variables with machine learning to downscale groundwater storage anomalies (GWSA) from 1° × 1° to 0.25° × 0.25° in the Breede Water Management Area, South Africa (2002–2022). The Random Forest model outperformed other tested models, revealing significant spatial heterogeneity in GWSA and accurately capturing major drought and recovery phases, thereby enhancing groundwater monitoring in data-scarce regions.

Objective

• To downscale GRACE-derived groundwater storage anomalies (GWSA) from a coarse spatial resolution of 1° × 1° to a finer resolution of 0.25° × 0.25° in the Breede Water Management Area, South Africa, using machine learning models.
• To evaluate the performance of Random Forest (RF) against Support Vector Regression (SVR) and Artificial Neural Networks (ANN) for this downscaling task.
• To analyze the spatial and temporal patterns of downscaled GWSA and validate them against in situ borehole observations to inform sustainable water resource management.

Study Configuration

• Spatial Scale: Breede Water Management Area, South Africa (approximately 12,600 km²). GRACE data at 1° × 1° (native 3°, mascon 0.5°), GLDAS data at 0.25° × 0.25°. Downscaled GWSA at 0.25° × 0.25°.
• Temporal Scale: 2002–2022 (21 years), monthly time steps.

Methodology and Data

• Models used:
  • Machine Learning: Random Forest (RF), Support Vector Regression (SVR), Artificial Neural Networks (ANN).
  • Land Surface Model: GLDAS Noah 2.1.
• Data sources:
  • Satellite: GRACE Jet Propulsion Laboratory (JPL) RL06 mass concentration (mascon) dataset.
  • Observation: In situ groundwater level data from the Department of Water and Sanitation (DWS) from three long-term monitoring boreholes.
  • Reanalysis/Model-derived: Global Land Data Assimilation System (GLDAS) Noah 2.1 for hydrological variables (canopy surface water, evapotranspiration, base flow groundwater runoff, root zone soil moisture, precipitation, temperature, soil moisture).

Main Results

• The Random Forest (RF) model consistently outperformed Support Vector Regression (SVR) and Artificial Neural Networks (ANN) in reproducing GWSA variability, achieving an R² of 0.86, a Root Mean Square Error (RMSE) of 0.0019 m, and a Nash–Sutcliffe Efficiency (NSE) of 0.75.
• Precipitation was identified as the most influential predictor for GWSA (Variable Importance Measure > 30%), followed by soil moisture and root zone soil moisture.
• The RF-downscaled GWSA (0.25° × 0.25°) revealed significant spatial heterogeneity, with higher groundwater recharge in the eastern mountainous regions and persistent depletion in the western lowlands.
• Temporal analysis captured major drought periods (2002–2006, 2015–2021) and recovery phases (2008–2010, 2020–2022), consistent with observed hydroclimatic variability.
• Validation against in situ boreholes showed strong agreement, with Pearson correlation coefficients ranging from 0.77 to 0.81, RMSE between 0.019 m and 0.025 m, and NSE scores up to 0.55.
• An overall net decline of -0.014 m in GWSA was observed from December 2007 to July 2020. Post-drought recovery in 2023 showed GWSA values ranging from 0.15 m to 0.40 m, but not fully returned to pre-drought levels in key areas.

Contributions

• Provides one of the first regional-scale applications of Random Forest for GRACE downscaling in the complex hydrogeological and climatic conditions of South Africa.
• Systematically evaluates and compares the effectiveness of RF, SVR, and ANN for GRACE downscaling, demonstrating RF's superior performance.
• Enhances groundwater monitoring in data-scarce regions by integrating satellite gravimetry with machine learning, offering higher spatial resolution GWSA maps.
• Generates actionable, high-resolution GWSA maps that can serve as a valuable decision-support tool for water managers to identify high-risk areas, monitor recovery, prioritize recharge zones, and inform drought preparedness and sustainable water allocation policies.

Funding

• University of the Witwatersrand (Open access funding).
• Pan-African Planetary and Space Science Network (PAPSSN) (Ph.D. bursary for the first author).

Citation

@article{Adam2026Application,
  author = {Adam, Maal A. and Scheiber-Enslin, Stephanie and Ali, K. A.},
  title = {Application of random forest modeling to evaluate groundwater storage changes in the Breede Water Management Area, South Africa},
  journal = {Hydrogeology Journal},
  year = {2026},
  doi = {10.1007/s10040-026-03047-w},
  url = {https://doi.org/10.1007/s10040-026-03047-w}
}