Estimating water levels in reservoirs using Sentinel-2 derived time series of surface water areas: a case study of 20 reservoirs in Burkina Faso

Codjia, Audrey Kantz Dossou; Akpoti, Komlavi; Dembélé, Moctar; Yonaba, R.; Fowe, T.; Sankande, S.; Koissi, Modeste G. Déo-Gratias; Zwart, Sander J. 2025. Estimating water levels in reservoirs using Sentinel-2 derived time series of surface water areas: a case study of 20 reservoirs in Burkina Faso. International Journal of Applied Earth Observation and Geoinformation, 139:104523. (Special issue: African EO Challenges and Solutions in Water Scarcity and Food Security – Mapping the Environment) [doi: https://doi.org/10.1016/j.jag.2025.104523]

Reservoirs play a significant role in the mobilization of water resources in Burkina Faso, contributing to the management and availability of water for various purposes. Operational management of reservoirs requires accurate and timely water level information, which remote sensing can provide cost-effectively and with limited resources. In this study, the surface area of 20 reservoirs is first determined using a Random Forest classifier and Sentinel-2 images acquired between 2015 and 2022. The accuracy of the classified surface water areas is evaluated by calculating 5 accuracy assessment metrics. The classifications were validated using manually digitized water areas from high-resolution Google Earth images and compared to the Dynamic World (DW) land cover dataset. Afterward, the spatial variation in the areal extent of the reservoirs is analyzed over time. A linear relationship is established between the estimated surface area and the corresponding observed water level of the reservoirs. The results indicate that reservoir surface areas were accurately classified with Sentinel-2 images (Kappa above 90.35%) for all dates. Moreover, validation with high-resolution images provided an R2 of 0.99 and a Normalized Root Mean Square Error (NRMSE) of 3.53%. Smaller reservoirs exhibit significant variations in surface areas over time as compared to larger ones, which are more stable. The relationship between surface area and water level is satisfactory (R2 ranging from 0.76 to 0.97) for 14 of the 20 analyzed reservoirs. The remaining six reservoirs are affected by aquatic plant intrusion which leads to an underestimation of the surface area. The high accuracy and operational feasibility of the proposed approach demonstrate that Sentinel-2 imagery and machine learning techniques can be recommended for reservoir mapping within the framework of water level monitoring in Burkina Faso.