Unlocking precision in flash flood forecasting: A synergistic approach using machine learning and satellite-ground integration.

Abstract

This research introduces an innovative approach to flash flood forecasting in poorly monitored or precipitation ungauged catchments. Leveraging data from satellite-based (PERSIANN-CCS) and ground-based precipitation estimates, our methodology is anchored in the Random Forest (RF) algorithm. We propose a hydrological model that integrates process-based Feature Engineering (FE) strategies, including a comprehensive connected component analysis (CCA) applied to satellite imagery, a soil moisture proxy, and subflow division. These strategies collectively aim to enhance the precision of short-term forecasting. We proved the utility of the proposed model in a 300-km2 catchment representative of the tropical Andes of Ecuador. While acknowledging potential reservations about the quality of hourly satellite data through an evaluation of the PERSIANN-CCS database at a microcatchment-wide scale, the proposed hydrological model achieved satisfactory forecasting models. We demonstrated forecasting performances for lead times up to three times thecatchment’s concentration time (Nash-Sutcliffe efficiency varying from 0.95 to 0.61). The success of our forecasting model is attributed to the synergistic fusion of ground- and satellite-based precipitation, coupled with FE strategies that integrate crucial physical knowledge into the models. Notably, the CCA enriches the input feature space of the RF models by deriving key precipitation attributes from precipitation objects otherwise hidden in timeseries. We assert the potential to exploit freely-available satellite information as an advantageous opportunity for the RF algorithm and other Machine Learning (ML) techniques. This refined approach not only facilitates efficient forecasting of peak flows w but also holds consequential implications for operational applications, including Flood Early Warning Systems (FEWSs), and contributes valuable insights into understanding precipitation-runoff responses in catchments constrained by sparse monitoring networks. Furthermore, the study unveils variations in the hyperparameterization of RF models across lead times, emphasizing the importance of incorporating physical knowledge into the modeling process. The inclusion of such insights, coupled with a meticulous approach to hyperparameterization, addresses persistent skepticism surrounding black-box, data-driven models. By highlighting ML hydrological models as a metaphorical "blank page," we advocate for their role as a testing ground where hydrological forecasting hypotheses can be scrutinized atop statistical and computational advantages.