Source: Geophysical Research Letters
Real-time hydrologic forecasting predicts river level and flooding inundation by combining continuously updated rainfall measurements, river gauge readings, and weather forecasts. Most of these flood forecasting systems depend on human interpretation and adjustments, or a “forecasters-in-the-loop” approach, which pairs computer models with a human expert on flood dynamics and local conditions. In contrast, in a “forecasters-over-the-loop” system, humans supervise automated forecasts and intervene only if necessary.
Recently, artificial intelligence (AI) and machine learning (ML) have become more integrated into flood prediction, and many of these systems are faster at processing large datasets and learning complex patterns from historical records than traditional models alone. But these new technologies also come with limitations—AI and ML require extensive data and may struggle to capture extreme, rare events.
Even though ML and AI are often touted as the future of flood forecasting, most studies have tested this technology against models that provide historical simulations, not the real-time operational systems that would be used during a flood. These simplified models may lack local details or are tested at daily rather than hourly resolution. Their effectiveness may be overestimated.
Tran et al. produced the first study comparing the performance of ML models to an actual flood forecasting system used at the California Nevada River Forecast Center (CNRFC) that uses professional forecasters and traditional hydrologic models. The study suggests that a forecasters-in-the-loop approach outperforms the ML models in several key ways, including streamflow predictions and flood event detection, because forecasters can recognize model errors and account for poor input data—actions models cannot take on their own.
The researchers used data gathered from CNRFC river stage forecasts across 50 California and Nevada locations between 2012 and 2022 and river condition lead times from 1 to 96 hours. Compared to the ML models, the Community Hydrologic Prediction System used at CNRFC generally performed better at predicting stream flow and flood peaks, especially with longer lead times. Though the ML models could perform better at very short lead times, their accuracy declined quickly. Though automated forecasting options may seem promising, they are not yet a suitable replacement for human expertise when it comes to protecting lives and livelihoods from damaging floods, the researchers say. (Geophysical Research Letters, https://doi.org/10.1029/2025GL118317, 2026)
—Rebecca Owen (@beccapox.bsky.social), Science Writer
Citation: Owen. R. (2026), Keeping humans in the loop improves flood forecasting, Eos, 107, https://doi.org/10.1029/2026EO260161. Published on 19 May 2026.
Text © 2026. AGU. CC BY-NC-ND 3.0
Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.
