Reading Shallow Flood Depth from Real USGS Gauge-Camera Footage with a Learned Staff-Plate Reader

1.13 cm
mean absolute error, held-out months, frames that pass the quality gate
0.79 cm
same, on two locked test months
67%
of frames pass the gate; the reader abstains on the rest
5.0 cm/px
resolution at which error is still under twice the native error
The reader on one held-out frame: frame with plate outline, rectified strip, row probabilities and decoded level, and error against image sharpness
The reader on one held-out frame. (a) Frame and plate outline. (b) Rectified strip. (c) Per-row probability of being submerged, decoded level (red) and gauge height (dashed). (d) Cross-validated error against image sharpness; green points pass the quality gate.

Abstract

How accurately can a fixed camera estimate flood water depth, and how well does that transfer across cameras? I use real USGS staff-plate cameras whose official gauge heights serve as labels. A learned reader registers each frame, rectifies the plate, classifies plate rows as submerged and decodes the water level. An image-quality gate abstains on bad frames.

On the main shallow-flow camera (2259 frames, 15 months), leaving out whole months gives 1.13 cm mean absolute error on the 67% of frames passing the gate (6.97 cm for a colour-threshold baseline), and 0.79 cm on two locked months. Error stays below twice its native value up to 5.0 cm per pixel.

Applied unchanged to three cameras downloaded afterwards, with each camera held out of training in turn, the reader beats a no-water guess on two cameras and fails on one. A river camera is read to 12 cm over a 211 cm flood.

Headline numbers

Mean absolute error in cm on frames that pass the quality gate. Every value on this page is generated from the files in results/ (see Reproduce).

EvaluationFramesLearned readerColour-threshold baselineAlways 0 cm
Leave-one-month-out CV (13 months) 1319 1.13 [0.67, 1.69] 6.97 8.8
Locked test months (2025-04, 2026-04) 233 0.79 11.52 18.9

Brackets are 95% intervals from a bootstrap over whole months. The two locked months are too few to give an interval. The gate drops 33% of CV frames and 18% of locked frames; without it the errors are 3.38 cm (CV) and 3.11 cm (locked). See Results for the full breakdown.

Where to go next

   
Method Registration, plate rectification, the row classifier, the decode step and the error model
Data The USGS cameras, where the gauge labels come from, and how the splits keep months apart
Results Held-out and locked-test errors, resolution sweep, other cameras, failure gallery
Demo Step through a real flood event frame by frame, in your browser
Paper The four-page PDF and a BibTeX entry
Reproduce Setup, make targets, and which script produces which result
Limitations What these results do and do not show