TORRENT dataset unites scattered flash-flood records across the United States

Flash floods are among the deadliest and most unpredictable hazards in hydrology. They develop within minutes to hours, strike mainly small to medium-sized watersheds, and often outrun the very instruments meant to capture them. Now a research team has unveiled TORRENT, a benchmark-ready community dataset that stitches together decades of scattered flash-flood evidence across the contiguous United States into a single, machine-readable resource designed for both human researchers and artificial intelligence tools.

The dataset, described in a preprint under review for the journal Earth System Science Data, was assembled by Songkun Yan, Mengye Chen, and colleagues at the University of Oklahoma, working with partners at NOAA’s National Severe Storms Laboratory, Vienna University of Technology, the University of Bristol, Columbia University, the University of Colorado Boulder, the University of Florida, and Tianjin Normal University. TORRENT stands for the Traceable, Observation-constrained, Rapid-Response, Episode–gauge–watershed Network of Testbeds, and its name captures its central ambition: to make flash-flood hydrological response traceable from the moment a flood is reported to the gauge measurement that confirms it.

The problem TORRENT addresses is one of fragmentation. Evidence of flash flooding has long been scattered across news reports, social media posts, stream gauges, operational storm event databases, and post-flood field surveys. Each source carries its own biases, formats, and gaps, which has historically limited advances in flash-flood research and made rigorous model evaluation or intercomparison difficult. Without a common benchmark, two rainfall–runoff models evaluated on different subsets of events cannot be meaningfully compared, and claims of improved flood prediction often rest on shaky ground.

TORRENT resolves this by linking three elements into what the team calls a testbed: one reported NOAA/NCEI flash-flood episode, one U.S. Geological Survey stream gauge, and the contributing watershed upstream of that gauge. Each testbed combines the observed discharge hydrograph—the continuous record of river flow that serves as the primary benchmark target—with watershed boundaries, temporally screened in-basin high-water marks, and quality-control metadata. Full packages additionally provide Multi-Radar Multi-Sensor (MRMS) precipitation forcing and auxiliary gridded products, giving modelers the rainfall inputs needed to drive simulations.

The dataset follows a deliberate three-level structure that distinguishes what was reported, what observations are available, and what can be evaluated reproducibly. Level 1 is a national inventory of 42,466 reported flash-flood episodes spanning 1996 to 2025. Level 2 documents observation availability for 5,424 recent episodes in the contiguous United States between 2021 and 2025, explicitly including episodes without usable external observations—a transparency measure that prevents silent selection bias. Level 3 contains 806 episode–gauge–watershed evaluation packages representing 160 episodes and 208 gauges, each with a retrievable watershed, at least one in-basin high-water mark meeting the stated temporal-screening criterion, and a detected discharge response.

At the top of this pyramid sits a conservative primary benchmark of 21 testbeds, distilled from 14 episodes and 21 gauges through a gauntlet of screening stages: flash-flood-response screening, deduplication, in-basin NCEI corroboration, station-relative severity screening, and post-selection curation. The authors are explicit about what this primary set is and is not. It is intended as a high-confidence reference for formal rainfall–runoff model intercomparison, not as a statistically representative sample of flash flooding across the contiguous United States. The broader Level-3 pool of 806 packages supports exploratory analysis and robustness testing where representativeness matters less than volume.

The primary testbeds were chosen for their rapid, pronounced hydrological responses, with estimated event-peak return periods ranging from about 2 years to more than 100 years. That range matters for model evaluation: a model that performs well on a 2-year flood may fail catastrophically on a 100-year event, and benchmark sets that exclude rare extremes can mask such failures. By anchoring each benchmark case to a gauge-measured discharge hydrograph rather than to a report alone, TORRENT ensures that model predictions can be scored against actual water moving through a channel, not against human impressions of severity.

Reproducibility is engineered into the dataset’s construction. The automated construction and screening stages are released as deterministic, re-runnable code, meaning any researcher can regenerate the dataset from raw inputs and obtain identical results. Documented curation decisions, sensitivity tests, and quality-control diagnostics accompany the release, so the judgment calls that shaped the benchmark are visible rather than buried. TORRENT is openly available on Zenodo under a Creative Commons Attribution 4.0 license, and its machine-readable design is aimed squarely at the growing community of AI and machine-learning flood models that require large, consistent training and evaluation corpora.

The implications reach beyond the United States. Flash-flood research has long been hampered by the same fragmentation problem worldwide, and a benchmark-ready template that links reports, gauges, watersheds, high-water marks, and radar-based rainfall could be adapted to other regions with dense observational networks. As climate change intensifies extreme rainfall and urbanization expands impervious surfaces into flood-prone terrain, the demand for trustworthy flash-flood prediction is rising. TORRENT gives the research community a common yardstick—and, for the first time, a national-scale library of flash-flood episodes in which every case can be traced, reproduced, and compared.

Subject of Research: A benchmark-ready community dataset linking flash-flood reports, stream gauges, and watersheds across the contiguous United States for model intercomparison

Article Title: TORRENT: An observation-constrained and benchmark-ready community shared dataset in the contiguous United States for flash-flood research and model intercomparison

Article References: Yan, S., Chen, M., Gourley, J. J., Zhang, J., Blöschl, G., Bates, P., Li, Z., Zhu, S., Snooks, A., Wen, Y., Chen, X., Deng, C., Toro, J. C. P., Ma, M., & Hong, Y. (2026). TORRENT: An observation-constrained and benchmark-ready community shared dataset in the contiguous United States for flash-flood research and model intercomparison. https://doi.org/10.5194/essd-2026-695

Image Credits: AI Generated

DOI: 10.5194/essd-2026-695

Keywords: flash floods, TORRENT dataset, hydrology, benchmark dataset, stream gauges, watersheds, rainfall-runoff models, model intercomparison, NOAA, USGS, machine learning, flood prediction