Introducing a new clustering-based method for regionalization framework for continental-scale rainfall estimates from soil moisture dynamics using machine learning methods

Abstract

Rainfall estimation plays a key role in various hydrological applications, ranging from flood forecasting and drought monitoring to water resource management. Traditional methods, which depend on ground-based gauges and remote-sensing products, can be expensive and limited by geography, and they often suffer from issues like sensor resolution or atmospheric interference. To tackle these problems, “bottom-up” strategies have emerged that use soil moisture as a stand-in for rainfall. By leveraging soil’s natural capacity to capture precipitation, these methods can reduce the reliance on high-resolution sensors and intricate modeling.

Nonetheless, their performance still depends heavily on careful calibration, a process that usually calls for plenty of on-site data, extended observation periods, or location-specific fine-tuning. To address these hurdles, we present a calibration parameters regionalization framework that does away with the need for a dedicated calibration phase. This framework uses both unsupervised (K-means clustering) and supervised (rainfall-intensity classification) techniques together with a genetic algorithm to automatically determine model parameters, without depending on adjustments tailored to specific regions.

We illustrate our method using the soil moisture to rainfall (SM2RAIN)-Net Water Flux (NWF) algorithm, demonstrating its ability to accurately estimate rainfall across the well-monitored contiguous United States (CONUS). Our findings indicate that SM2RAINNWF performs particularly well in areas with higher rainfall intensity, outperforming the classic SM2RAIN methods that are commonly used for estimating rainfall from soil moisture dynamics. In fact, this is the first time K-means, a genetic algorithm, and rainfall clustering have been combined to estimate rainfall without requiring a separate calibration period, achieving a 20 % improvement in Nash–Sutcliffe efficiency and a 10 % reduction in root mean square error compared to classical methods.