Monitoring and Predicting Hydrological Extremes Using Remote Sensing and Machine Learning 3 views

Freshwater systems worldwide face a growing observation gap: hydrological variability is intensifying due to climate change and land transformation, yet many large river basins remain effectively ungauged at scales relevant to management. This dissertation develops integrated remote sensing and machine learning frameworks to recover hydrological information in data-scarce environments, enabling retrospective diagnosis and operational prediction of hydrologic extremes across contrasting hydroclimatic settings. This work is anchored in South America's La Plata River Basin (LPRB), which sustains agriculture, hydropower, and water supply for over 100 million people. The first study characterized the severe 2019–2021 drought using GLDAS and SMAP satellite observations, with Mann-Kendall trend analysis revealing that water loss concentrated in the Upper Paraná subbasin (UPRB)—reframing a perceived basin-wide deficit as a regionally focused hydrologic collapse. Building on this diagnosis, the second study evaluated whether GRACE terrestrial water storage anomalies (TWSA), combined with climate variables, can reconstruct reservoir water levels across 14 major UPRB reservoirs without in situ records. The third study investigated drivers of long-term storage change in the Itaipu hydropower region, identifying a significant basin-wide TWSA regime shift in mid-2009 coinciding with cropland expansion. Finally, the framework is applied to flash flood prediction in Ireland's Lee River Basin across 134 storm and flood events. Using a dual-experiment design to prevent data leakage, satellite-derived ERA5-Land precipitation with an Extra Trees classifier achieved pre-event magnitude classification 63% above the random baseline, matching gauge-based skill in an ungauged setting. Together, these studies demonstrate that satellite observations and machine learning can redefine hydrological monitoring where traditional infrastructure falls short, offering a scalable pathway for drought detection, flood forecasting, and attribution of hydrologic change.

PHD (Doctor of Philosophy)

Remote Sensing; Hydrology; Hydrological Extremes; Droughts; Floods; Machine Learning

English

2026-04-21