Demo
Everything here runs in your browser from small precomputed files; nothing is sent anywhere.
The depth readings in the first two panels are held-out estimates: each one comes from a model trained without that month (September 2026), so the reader had never seen these frames. The reading is not computed live in the page. The page replays stored results next to the official USGS gauge height.
1. Frame explorer: a real event, frame by frame
Camera in_overland, 17–22 September 2026, every third frame. The yellow outline on the photo is the staff plate found by registration; the strip on the right is that plate rectified to the reader's input. Drag the slider or press play.
Lines on the strip use the plate's nominal geometry (ruler marks in cm). The reader also applies a fitted affine correction and the gauge sits beside the plate, so a line can sit a few cm from the visible waterline. Frames are the 640 px public copies, so the strip is softer than the 1280 px input the reader normally gets. Some frames are rejected on purpose: night, fog and a corrupted image are all in this event.
2. Resolution: what happens when each pixel covers more centimetres
The strip of the frame selected above, coarsened by a factor f. The error table is from the study, where the reader was retrained at every factor.
Illustration only: the strip is area-averaged in your browser from the native strip. The study downsampled the whole frame before re-sampling the plate, so it is close but not identical. The numbers come from results/real/tables/resolution.csv. The error model is \(\mathrm{RMSE}^2 = s^2\sigma_v^2 + \sigma_g^2\) with \(\sigma_v = 0.55\) px and \(\sigma_g = 3.0\) cm.
3. State machine playground (dashboard logic, not an evaluated result)
A JavaScript port of the dashboard's rolling-median filter and four-state machine (Normal, Pending, Flood, Recovering) with a 10.5 cm trigger and a 9.5 cm clear level. Shape a synthetic flood and watch the state change. There is no recorded data here.
The Recovering state is a dashboard demonstration. Short clips cannot validate trends that unfold over minutes to hours, so draining versus accumulating is not an evaluated claim of this project.