Reconstructing reservoir water levels using basin-scale GRACE TWS and machine learning: A case study in the Upper Parana River Basin

We examine 14 large reservoirs (115–1156 km²) in the Upper Paraná River Basin, Brazil, where reservoir water levels are shaped by hydrological variability and operational water management decisions. Effective monitoring in this basin is challenged by limited in-situ observations, creating a need for alternative approaches in data-scarce environments. This study evaluates whether satellite-derived Terrestrial Water Storage (TWS) anomalies from GRACE, combined with precipitation, temperature, and reservoir-specific attributes, can accurately reconstruct reservoir height anomalies across this diverse reservoir system. Linear Regression, Polynomial Regression, Random Forest, and Long Short-Term Memory networks were trained and evaluated using multi-satellite radar altimetry observations as the target. In addition, feature importance analysis was conducted to identify the dominant inputs driving height variability. Results reveal a trade-off between system-wide stability and reservoir-specific accuracy. Polynomial Regression achieved the most consistent system-wide (MAE = 1.50 m), while Random Forest excelled at individual sites (MAE < 1.0 m at 9 out of 14 reservoirs), revealing a fundamental trade-off between stability and site-specific optimization. TWS emerged as the most influential input, underscoring the importance of integrating basin-scale storage dynamics with site-level characteristics. This work demonstrates a scalable framework for reconstructing reservoir water levels, improving the understanding of historical reservoir behavior, and supporting flood mitigation and drought preparedness in data-limited regions. • Reservoir water level reconstruction using machine learning algorithms. • Remote sensing of reservoirs using multi-satellite radar altimetry and GRACE data. • Improvement of reservoir level reconstruction with meteorological datasets.