Streamflow patterns across the north-central United States are shifting in ways that make flooding harder to anticipate, according to a peer-reviewed U.S. Geological Survey study spanning nine states and nearly 75 years of gage records. At the same time, federal researchers have begun releasing machine learning tools designed to forecast weekly river conditions up to 13 weeks ahead, offering a potential antidote to the growing unpredictability. The collision of these two developments, one documenting a worsening problem and the other proposing a technological fix, frames a critical question for prairie communities that depend on accurate water forecasts for agriculture, infrastructure, and emergency planning.
Flood Timing Is Drifting Across Nine States
The USGS study analyzed data from hundreds of unregulated streamgages across Minnesota, North Dakota, South Dakota, and six other north-central states over the period 1946 to 2020. Researchers examined trend windows of 75, 50, and 30 years through water year 2020, looking for measurable changes in when peak flows arrive and how concentrated seasonal runoff has become. The shorter windows revealed the most pronounced shifts, suggesting that the pace of change is accelerating rather than holding steady. The resulting assessment of changing streamflow seasonality concludes that many rivers are now hitting their highest flows earlier in the year than they did in the mid-20th century.
What makes this work unusual is its focus on timing rather than just volume. Most flood research asks how much water a river will carry; this analysis asks when that water will show up. The distinction matters because emergency managers, dam operators, and farmers all calibrate their preparations around expected timing. When spring melt arrives weeks earlier than historical norms, or when summer storms generate peak flows that once belonged to a different season, the old playbook stops working. Earlier peaks can collide with frozen ground or lingering snowpack, while later, rain-driven peaks can threaten crops and infrastructure that were not previously at risk during those months.
For local governments, that drift in seasonality complicates everything from road design to reservoir rule curves. Communities that once counted on a relatively narrow “flood season” now face a longer window of elevated risk. Insurance pricing, evacuation planning, and even planting schedules depend on assumptions about when rivers are most dangerous. As those assumptions erode, the value of long historical records lies less in predicting exact dates and more in revealing how quickly patterns are changing.
Why Prairie Watersheds Are Especially Hard to Read
Prairie hydrology does not behave like mountain or coastal hydrology. Flat terrain, extensive wetland networks, and highly variable soil moisture create a system where small changes in precipitation or temperature can produce outsized swings in runoff. As hydrologist Ali Ameli explained in a recent analysis of prairie conditions, the region has experienced larger swings between very wet and very dry years, with consequences for both water availability and water quality. That volatility compounds the forecasting challenge because wetlands act as sponges up to a threshold, then suddenly shift from storing water to releasing it once saturated.
Conventional physics-based models struggle with this behavior. NOAA’s National Water Model, which runs operational forecasts across the contiguous United States, Alaska, Hawaii, and Puerto Rico/U.S. Virgin Islands, provides high-flow magnitude guidance based on a multi-decade reanalysis. That system is powerful at continental scale, but prairie-specific dynamics, especially the role of wetland connectivity and frozen soils, remain difficult to capture in a single national framework. Even with historic simulations restructured for efficient time-series access, modelers still face gaps in representing how thousands of small depressions, tile drains, and temporary channels interact during extreme events.
These complexities mean that two storms with similar rainfall totals can produce very different outcomes depending on antecedent wetness and the degree of wetland fill. In a dry year, much of the water may disappear into soils and storage basins; in a wet year, the same storm can generate rapid, basin-wide runoff. For farmers and towns scattered across the Dakotas and Minnesota, that difference can be the line between a manageable high flow and a damaging flood.
Machine Learning Models That Look Weeks Ahead
Against this backdrop, USGS researchers have released both model code and outputs for machine learning systems that forecast weekly streamflow percentiles at more than 3,000 gage locations across the country for lead times of 1 to 13 weeks. The models, described in a dataset of streamflow and drought forecasts, rely on Long Short-Term Memory networks (LSTM) and LightGBM algorithms and are benchmarked against simpler persistence and ARIMA approaches. By making the code open-source, the agency has invited independent testing and adaptation, a step that matters because trust in AI-generated forecasts depends on transparency and reproducibility.
These tools do not replace short-term flood warnings issued days in advance, but they aim to answer a different question: how likely is it that a given river will be unusually high or low several weeks from now? That kind of outlook can help reservoir managers decide when to release water, guide utilities planning for hydropower output, and inform emergency managers about when to pre-position resources. In agricultural regions, multi-week streamflow guidance can also signal when irrigation withdrawals may strain ecosystems or when saturated conditions could delay planting.
Ameli argued that AI can support flood preparedness by helping identify when a watershed is getting close to the point where wetlands begin to dominate downstream flows. That framing highlights a specific advantage of data-driven methods: they can detect nonlinear thresholds in historical records without requiring a full physical description of every wetland, drainage ditch, and soil layer in a basin. In practice, that might mean identifying combinations of snowpack, soil moisture, and forecast precipitation that have historically preceded rapid rises, even when traditional models fail to capture the underlying mechanics.
New Research Challenges Old Forecasting Assumptions
Separate work from the University of Minnesota Twin Cities, summarized in a recent news release, found that a hybrid machine learning approach can predict streamflow and flood levels with greater accuracy than methods currently used across much of the United States. The university’s team, whose findings are discussed in coverage of increasingly unpredictable prairie watersheds, combined machine learning with process-based insights to capture both statistical patterns and hydrologic mechanisms. Their results suggest that AI does not have to be a black box: when paired with physical understanding, it can complement rather than replace traditional models.
Together, the USGS and university efforts challenge long-standing assumptions about how far into the future river conditions can be forecast with useful skill. Instead of treating weekly to seasonal horizons as a blind spot between weather forecasts and long-term climate projections, researchers are carving out a middle ground where probabilistic streamflow outlooks become part of routine planning. For prairie communities facing both more volatile precipitation and shifting flood seasons, that added lead time could mean the difference between reactive sandbagging and proactive mitigation.
From Research to Practice on the Prairies
Turning these advances into real-world benefits will require more than code repositories and journal articles. Local agencies need training and guidance on how to interpret probabilistic forecasts, how to integrate AI outputs with existing warning systems, and how to communicate uncertainty to residents. In many rural areas, staffing and budgets are limited, making it difficult to evaluate new tools while managing day-to-day responsibilities.
Federal resources can help bridge that gap. The U.S. Geological Survey maintains an online catalog of maps and data products that includes streamflow records, flood-inundation maps, and other decision-support materials. For practitioners and residents with specific questions about how to access or apply these resources, the agency’s public information portal offers a way to connect with subject-matter experts. As AI-based forecasts mature, similar channels may become crucial for explaining model strengths and limitations in plain language.
On the ground, the next phase of work will likely focus on pilot projects that pair prairie communities with researchers to co-design applications. That could mean tailoring 13-week forecasts to support wetland restoration planning, testing hybrid models in basins with recent flood damage, or integrating machine learning outputs into state-level drought and flood dashboards. Each experiment will provide feedback on where AI adds value, where it falls short, and how best to combine it with local knowledge.
The underlying message from recent studies is not that technology can eliminate uncertainty. Instead, the research underscores that as climate variability reshapes streamflow timing, clinging to historical patterns is no longer enough. By blending long-term observations, physics-based models, and machine learning, prairie regions may be able to navigate a future where water arrives at unfamiliar times, without being caught entirely off guard.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
