Downscaling GRACE Total Water Storage Data Using Random Forest: A Three-Round Validation Approach under Drought Conditions

Youssef Hamou-Ali¹Nourlhouda Karmouda¹Ismail Mohsine¹Tarik Bouramtane^1*ILIAS KACIMI¹Sarah Tweed^1,2Mounia Tahiri¹Nadia Kassou¹ALI El Bilali³Omar Chafki³Mohamed Abdellah Ezzaouini³Siham Laraichi³Abdelaaziz Zerouali³Marc Leblanc^1,2,4*

• ¹Faculty of Science, Mohammed V University, Rabat, Morocco
• ²Institut de Recherche Pour le Développement (IRD), Marseille, Provence-Alpes-Côte d'Azur, France
• ³River Basin Agency of Bouregreg and Chaouia, Chaouia, Morocco
• ⁴University of Avignon, Avignon, Provence-Alpes-Côte d'Azur, France

The application of Total Water Storage (TWS) data from GRACE satellites to local water management is constrained by its coarse spatial resolution (100-300 km). To address this, Random Forest-based model was employed to downscale GRACE TWS data from 100 km to 1 km over Morocco, a region severely affected by drought, covering the period from 2002 to 2022. Input data included precipitation (GPM, 10 km), NDVI (MODIS, 1 km), land surface temperature (LST, MODIS, 1 km), evapotranspiration (MODIS, 500 m), elevation (SRTM, 30 m), and the normalized difference snow index (NDSI, MODIS, 500 m). Downscaling GRACE data improves spatial resolution, but validating these finer-scale results is a challenge. In this study, downscaled data were validated using three approaches: Statistical Validation , Validation using groundwater level in situ data , and Validation using known aquifer dynamics.Statistical validation demonstrated strong model performance, with a Nash-Sutcliffe Efficiency (NSE) of 0.80, a low RMSE of 0.82 cm, and an MAE of 0.57 cm, along with an R² of 0.80 between the original and downscaled data. Cross-validation confirmed the model's consistency, with mean, median, and maximum R² values of 0.56, 0.64, and 0.89, respectively. Error metrics remained low throughout the study period, with MAE values ranging from 0.36 cm to 0.6 cm and RMSE values between 0.5 cm and 0.8 cm. Additionally, the comparison with in-situ groundwater levels showed significant improvements, with correlation coefficients increasing for 63% of the 139 analyzed wells. The 1 km TWS data revealed localized variations and more distinct trends across different aquifers, while aquifer systems within the same structural domain tended to exhibit similar TWS patterns. These findings highlight the downscaling model's potential for improving local water management by capturing finer hydrological variations.