Check the landing page. If colored cards are showing, creeks are running — tap one or head to Gauges for the full picture. If it says "All Quiet," check the Watersheds page to see if rain is on the way. The site updates every 15 minutes, automatically.
This site pulls together everything a whitewater paddler in Arkansas, Oklahoma, and Missouri might need to make an informed go/no-go decision — all in one place, updated around the clock.
It monitors 25 creeks and rivers across the region using data from USGS stream gauges, NOAA radar rainfall observations, and National Weather Service precipitation forecasts. Instead of bouncing between government websites, weather apps, and group chats, you get a single dashboard that answers three questions:
For the Cossatot River, Richland Creek, and the Hailstone run on the Upper Buffalo, the site goes further — dedicated intelligence pages with physics-based prediction models and a nightly AI-driven calibration review that estimate when a creek will rise, how high it will get, and when it will cross paddling thresholds.
The backbone of the site is the USGS (United States Geological Survey) gauge network. These are physical instruments installed in creeks and rivers across the country. Each gauge reports measurements roughly every 15 minutes.
The site tracks two primary measurements:
One creek — Wister Wave — uses a US Army Corps of Engineers dam release gauge instead of a USGS gauge, since the wave depends on controlled dam releases rather than rainfall.
QPE stands for Quantitative Precipitation Estimate. It's radar-observed rainfall — a measurement of rain that has already fallen, derived from the national MRMS (Multi-Radar Multi-Sensor) system. Think of it as a satellite view of where rain hit the ground and how much accumulated.
The site uses MRMS data at roughly 1-kilometer resolution, updated every few minutes. This is what powers the Watershed alerts and the prediction models — it tells us exactly how much rain fell in the headwaters of each creek.
QPF stands for Quantitative Precipitation Forecast. This is the National Weather Service's prediction of how much rain will fall in the next 6 to 24 hours.
Important caveat: QPF is useful as a general heads-up that rain is coming, but it is unreliable for the kind of fast-moving thunderstorms that are common in the Ozarks. Convective storms are notoriously hard to predict in terms of exact location and intensity. Don't plan a trip solely on QPF — use it as a "be ready" signal, not a guarantee.
The site uses a consistent four-color system across all pages. Once you learn it, you can read any page at a glance.
These colors tell you whether a creek is paddleable based on the current gauge reading:
These colors tell you whether enough rain has fallen (or is falling) in a creek's headwaters to potentially produce a runnable rise:
These terms intentionally borrow from NWS severity conventions (WATCH = conditions are possible, WARNING = conditions are expected).
The "Recent Rain" column on the Watersheds page shows how much rain has fallen in the past 7 days, with colors indicating ground saturation:
On the Gauges page, trend arrows show which direction the water level is moving:
The landing page at creeks.druidnetworks.com is designed for quick mobile checks — glance at it from the truck at the put-in.
The Gauges page is the full detail view — a table with every creek the site monitors. Here's what each column means:
You can sort the table by level (default) or alphabetically by name. Tap the column headers to toggle.
The Watersheds page shows the rainfall picture — what's happening in the headwaters of each creek system.
Drainage: A geographic area in a creek's headwaters where we monitor rainfall. Rain falling in a drainage flows downhill into the creek. Each drainage has a defined center point and radius where we sample radar rainfall data.
Family: Drainages are grouped into families because multiple drainages often feed the same creek or creek system. For example, the Richland Family includes Upper Richland, Falling Water, Big Devils, Long Devils, and Bobtail — all of which drain into Richland Creek.
Trigger Threshold: The amount of rainfall (in inches) needed within a specific time window to produce a runnable rise. For example, "1.5 inches in 12 hours" means that if 1.5" of rain falls within a 12-hour window in that drainage area, the creek should come up. These values are calibrated against real events and adjusted for recent ground moisture.
Lag Time: How long after rain falls in the headwaters before the gauge downstream sees the rise. Large watersheds like the Upper Cossatot have 8-10 hour lags. Tiny micro-drainages like Adkins or Bobtail show "NOW" — the gauge is essentially in the drainage, so the response is nearly immediate.
Each row is one drainage. The columns show:
The site uses three progressively more sophisticated approaches to predicting when creeks will run.
This is the simplest and most broadly deployed approach, and it powers the WATCH/WARNING/FLOOD alerts on the Watersheds page.
The logic: "Did enough rain fall, in a short enough window, in the right geographic area?" Each drainage has a calibrated trigger — for example, Upper Cossatot needs 1.25" in 12 hours. The system checks radar rainfall against these triggers every 15 minutes and adjusts for ground moisture (dry ground needs more rain, wet ground needs less).
When a trigger is met, the alert holds for the lag time plus a 2-hour buffer, so the alert stays active while the pulse of water is still in transit to the gauge.
The Cossatot, Richland, and Hailstone intelligence pages each include a physics-based prediction model.
Instead of a simple "enough rain? yes/no" check, the physics predictor divides the watershed into travel-time bands (Cossatot's 5 bands from 0-2 hours to 8-10 hours) or precipitation zones (Richland's two HUC12 zones; Hailstone's single Boxley drainage). It tracks where in the watershed the rain fell and models how long water from each area takes to reach the gauge.
Think of it like watching a pulse of water move downstream. Rain that falls near the gauge arrives first. Rain that falls in the upper headwaters takes much longer. By knowing how much rain fell in each zone and how long it takes to arrive, the model predicts:
The model also accounts for ground moisture — wetter ground produces more runoff from the same amount of rain.
Every prediction the model makes is archived. The Physics Predictor Archives card on each intelligence page opens a per-day log of every hourly forecast, with the actual peak filled in after the peak window closes — useful for sanity-checking how often the model is hitting vs. missing.
After midnight each day on which there was meaningful rain or gauge response, a local AI model reviews:
The review writes a short markdown report (linked from each intelligence page as Nightly AI Analysis Archive) classifying the day's prediction (correct / partial / false_positive / false_negative), explaining what the model got right or wrong, and — when the evidence supports it — adjusting the physics model's coefficients within bounded constraints (±20% per event maximum) so tomorrow's prediction is slightly better calibrated. The empirical forecast is graded but not auto-tuned; its percentile tables are regenerated separately from longer historical record.
The AI runs against a local inference server (no external API calls), so no usage costs accrue and no data leaves the network.
An LSTM-based machine learning model used to publish hourly flow forecasts on the Cossatot and Richland pages alongside the physics predictor. It was retired in April 2026 after the physics predictor + nightly AI calibration loop produced more reliable per-event accuracy with less noise. The underlying inference still runs in the background for monitoring, but it no longer surfaces in the UI.
Current coverage: The physics predictor is available on the Cossatot, Richland, and Hailstone intelligence pages. The simpler watershed trigger system covers all other creeks. As the models prove out, additional creeks may get dedicated prediction pages.
When a creek is falling, the Gauges page and the Cossatot, Richland, and Hailstone intelligence pages show an estimated time until the creek drops below each level.
The system fits an exponential decay curve to the recent falling gauge readings — essentially, it measures the rate of decline and projects forward. Creek recessions tend to follow a predictable curve: fast at first, then gradually tapering off.
These dedicated dashboards provide a much richer picture than the main gauge table for these three runs. They're designed for paddlers who are actively tracking an event or deciding whether to make the drive. All three pages follow the same card layout so reading any one of them transfers cleanly to the others.
Each page includes (in order, top to bottom):
These pages auto-refresh every 5 minutes during active events.
The three pages share their underlying architecture but each is calibrated independently — Cossatot is height-based on a 5-band watershed, Richland is height-based on two HUC12 zones, Hailstone is CFS-based on the Boxley drainage (same physical gauge as Boxley but mapped to whitewater-floatable CFS tiers of 500/700/2000).
The Buffalo River Intelligence page (/buffalo/) is a read-only dashboard for the 7 gauges along the Buffalo (Boxley → Harriet, plus the Richland and Bear Creek tributaries): current flows, radar rainfall by sub-watershed, flood-risk and downstream propagation alerts, and a 3-day NWS forecast. Two features focus specifically on the Ponca gauge and the Ponca→Pruitt section that outfitters run.
The National Park Service classifies this section by flow and only permits commercial float launches between 100 and 1,600 cfs:
| Bucket | Flow (cfs) |
|---|---|
| Very Low | under 100 |
| Low | 100–200 |
| Moderate | 200–900 |
| High | 900–1,600 |
| Flood | 1,600 and up |
These buckets (not the generic creek tiers used elsewhere on the site) are what the Ponca historical report and the AI analysis card use.
Linked from the Buffalo page (below the 3-day forecast), the Ponca historical report is a deep look back at how this section has behaved:
When rain is actively falling over the Ponca headwaters (roughly a quarter-inch or more accumulated), an orange card appears at the top of the Buffalo page, above the gauges. It compares the unfolding storm — rain so far, current flow and trend, recent wetness — against the past storms in the historical catalog, and writes a short plain-English read of where things are likely headed: the most probable peak stage, and whether there's meaningful risk of climbing past the 1,600 cfs launch ceiling.
A few things to understand about it:
The Buffalo River Watershed Study is a 90-day research project (March through June 2026) designed to answer fundamental questions about how the Buffalo River responds to rainfall.
The study monitors 7 USGS gauges from the headwaters to the lower river (Boxley, Ponca, Pruitt, St. Joe, Richland, Bear Creek, and Harriet) and tracks radar rainfall across 37 sub-watersheds spanning the entire Buffalo River drainage.
The study is investigating four tiers of watershed response:
Every night, an AI system (Claude Opus) analyzes the day's gauge and rainfall data in the context of everything observed so far during the study. It produces a daily analysis report and updates a running hypothesis document that evolves as new storm events provide more data.
The study's daily analysis reports and the evolving hypothesis document are available on the Study pages for anyone interested in following along.
The long-term goal: the findings from this study will inform a Buffalo River prediction system similar to what exists for the Cossatot and Richland — but that work comes after the study concludes and the data has been analyzed.
No system is perfect. Here's what this site can and can't do:
This site is built for the paddling community and improves with community input.
Know of a USGS-gauged creek that should be on this site? Use the Suggestions page — Creek tab. You'll need the USGS gauge ID and your best estimate of paddling thresholds.
If the site says "Optimal" but your experience says otherwise, tell us. Use the Suggestions page — Correction tab. The more specific you can be (e.g., "Optimal should start at 3.0 ft, not 2.5 — I've run it at 2.8 and it's still too bony"), the faster we can calibrate.
If you know about a headwater area that feeds a creek and isn't currently monitored, use the Suggestions page — Drainage tab. Approximate lat/lon is fine — we can refine it.
What's useful? What's confusing? What would make this site more valuable to you? Use the Suggestions page or reach out to the Arkansas Canoe Club community.
Why does it say "All Quiet" when it's raining outside? The rain you see out your window may not be falling in the headwaters of any monitored creek, or it hasn't accumulated enough to cross a trigger threshold yet. Check the Watersheds page — you might see WATCH status developing.
Why is the Cossatot page so much more detailed than other creeks? The Cossatot was the first creek to get a dedicated prediction system with travel-time band analysis and machine learning forecasts. Richland Creek now has a similar setup. Other creeks use the simpler watershed trigger system. We're expanding the advanced prediction approach to additional creeks over time.
How often does the data update? Every 15 minutes for gauge readings and rainfall. The Cossatot and Richland intelligence pages auto-refresh every 5 minutes in your browser.
What does "stale data" mean? If a data source hasn't updated within its expected window (e.g., gauge data older than 30 minutes), the site shows a warning banner. The data is still displayed — it might just be a temporary USGS reporting delay.
Can I get notifications when a creek comes up? The site currently sends Signal alerts to a paddling group when watershed status escalates (e.g., from No Alert to WARNING). Contact the site operators if you'd like to be added to the notification group.
What are the AW and OZP links? AW is American Whitewater — the national nonprofit for whitewater paddling, with run descriptions, access info, and difficulty ratings. OZP is Ozark Paddler — a regional resource with detailed write-ups on Arkansas and Oklahoma runs.
| Term | Definition |
|---|---|
| AW | American Whitewater — national whitewater advocacy and information nonprofit |
| CFS | Cubic Feet per Second — a measure of water flow volume |
| Drainage | A geographic area in a creek's headwaters where rainfall is monitored |
| Height (ft) | Water level at a gauge, measured in feet above a fixed datum |
| HUC12 | A USGS sub-watershed boundary code, used to organize drainage areas |
| Lag Time | The delay between rain falling in headwaters and the gauge downstream responding |
| MRMS | Multi-Radar Multi-Sensor — NOAA's high-resolution radar rainfall system |
| Neural Net (LSTM) | A machine learning model trained on historical data to forecast streamflow |
| OZP | Ozark Paddler — regional whitewater information resource |
| QPE | Quantitative Precipitation Estimate — radar-observed rainfall (what already fell) |
| QPF | Quantitative Precipitation Forecast — predicted future rainfall |
| Recession | The falling limb of a creek's hydrograph after rain stops |
| Trigger Threshold | The rainfall amount needed in a time window to produce a runnable rise |
| USACE | US Army Corps of Engineers — operates dams and controlled-release water features |
| USGS | United States Geological Survey — operates the national stream gauge network |
| Watershed | The total land area that drains into a particular creek or river |
This site relies on public data from the USGS, NOAA, the National Weather Service, and the US Army Corps of Engineers. Creek information and run descriptions are linked from American Whitewater and Ozark Paddler. Paddling thresholds are calibrated with feedback from the Arkansas Canoe Club community and regional paddlers.
Built and operated as a community resource for the paddling community of Arkansas and surrounding states.
Last updated: April 2026