A moving pattern in a river video is evidence to interpret, not yet a discharge measurement
A river video can reveal motion without putting an instrument into the flow. But pixels per frame are not metres per second, and surface speed is not automatically depth-averaged speed. A defensible estimate follows a chain: identify the tracked pattern, establish geometry and timing, remove camera movement, estimate surface velocity, then justify any conversion to discharge.
This guide explains that chain through fictional numerical cases. It complements the existing acoustic and tracer-dilution guides rather than replacing them. The central question is what each stage actually adds to the evidence, and which uncertainty remains after a visually convincing velocity map appears.
This is an explanation of measurement reasoning, not instructions to fly an aircraft, enter a river, place tracers or assess a crossing. Real observations require appropriate access, aviation compliance where relevant, professional field planning and current site conditions. No worked value below is a recommended correction for a real river.
Hydrology subject guide · Learning that transfers into life · How X Works hub
Choose your reading route
Route 1: Define the measurement · Chapters 1–2
Route 2: Control geometry and time · Chapters 3–5
Route 3: Read motion responsibly · Chapters 6–8
Route 4: Calculate and test discharge · Chapters 9–10
Convert surface motion to a depth-averaged estimate cautiously
Route 5: Validate and transfer · Chapters 11–12
Full chapter index · Independent worked case · Questions and practical checks
Chapter index
Define the measurement · Chapters 1–2
Separate image motion, surface velocity and discharge
Know which image method produced the vector
Control geometry and time · Chapters 3–5
Convert pixels using the water-surface geometry
Use elapsed time rather than an assumed frame rate
Remove camera motion without removing river motion
Read motion responsibly · Chapters 6–8
Distinguish water-following features from waves and glare
Choose comparison regions and reject unsupported vectors
Use velocity normal to the measuring section
Calculate and test discharge · Chapters 9–10
Convert surface motion to a depth-averaged estimate cautiously
Integrate area and velocity without multiplying incompatible averages
Validate and transfer · Chapters 11–12
Validate the complete chain with genuinely new evidence
Report the result at the level the evidence supports
CHAPTER 1 OF 12
1. Separate image motion, surface velocity and discharge
A camera records changing light at positions in an image. Image velocimetry interprets displacement of recognisable patterns between observations. Large-scale particle image velocimetry, often shortened to LSPIV, commonly compares patterns within small image regions. The output initially describes apparent motion in image coordinates. Calling that output river velocity already assumes that the patterns represent moving water and that the camera geometry is understood.
Surface velocity describes movement at or near the water surface. Discharge is the volume passing through a cross-section per unit time, requiring velocity over the entire wetted area. A fast surface streak cannot establish the depth profile beneath it. The useful conceptual separation is observation, geometric conversion and hydraulic interpretation. Each stage can work correctly while the next remains unsupported.
Consider a clip showing a bright patch moving across ten pixels. Before calculating a speed, ask whether it is floating foam, reflected sunlight or a wave crest. Ask how long the displacement took and how many ground metres ten pixels represent there. Then ask what part of the cross-section the observation represents. These are necessary measurement questions, not optional refinements after the number has been produced.
CHAPTER 2 OF 12
2. Know which image method produced the vector
Pattern correlation and individual particle tracking do not make identical observations. A correlation method estimates a representative displacement within an interrogation region; particle tracking follows identifiable features through frames. Space-time image methods analyse motion along selected image lines through time. Their strengths and failure modes differ. Preserve the actual method rather than treating every coloured arrow on a video as interchangeable evidence.
A correlation peak can be sharp even when the tracked feature is wrong for the hydrological question. A repeating reflection or stationary texture can create an apparently stable signal. Conversely, real water motion may be difficult to estimate where the surface has little visible texture. Method confidence concerns the image calculation; hydrological validity also requires evidence that the selected features follow the flow sufficiently well.
For transfer, compare a textured, sediment-laden river with clear water reflecting a smooth sky. Keeping the software settings unchanged does not guarantee comparable performance. The second scene may contain less useful moving texture. A correct response may be to reject parts of the image or use another measurement method, rather than interpreting every output vector simply because the program can calculate one.
CHAPTER 3 OF 12
3. Convert pixels using the water-surface geometry
A simple overhead teaching image might have an approximately uniform scale of 0.025 m per pixel over a small level surface. A displacement of 16 pixels then represents 0.40 m. An oblique image usually has a scale that changes with position: distant pixels can represent more ground distance than nearby ones. One ruler placed near the camera is not enough to justify a uniform scale across the river.
Georeferencing or image rectification relates image positions to a defined spatial surface using suitable control information. Water level matters because the visible water surface can lie above or below the surface used in the transformation. Lens distortion, camera pose and uncertain control points can contribute additional error. The transformation should be appropriate to the geometry, not merely capable of producing a rectangular-looking image.
A changed-context check raises the river level while keeping a bank-mounted camera fixed. The camera has not moved, but the observed water surface has. Test whether the original mapping still represents distances on that surface. If a revised transformation changes inferred velocities, preserve both geometry versions and explain the reason. A geometric correction is a change in measurement interpretation, not a sudden acceleration of the river.
CHAPTER 4 OF 12
4. Use elapsed time rather than an assumed frame rate
Velocity is displacement divided by elapsed time. If 0.40 m of displacement occurs over 0.20 s, speed is 2.0 m/s. The time should come from reliable acquisition information and the frames actually compared. A displayed playback rate can differ from acquisition timing after conversion, editing, dropped frames or slow-motion processing. A visually smooth clip is not proof of a uniform physical sampling interval.
Suppose a video file plays at 30 frames per second but was acquired at 60 frames per second and deliberately slowed for presentation. Treating its playback timing as acquisition timing halves the inferred speed. The spatial displacement is unchanged. This example shows why a polished presentation copy should not silently replace the original measurement record. Preserve timestamps, frame selections and any processing that changes timing.
For a temporal-resolution check, compare adjacent frames with frames separated by a longer interval. Very short intervals may produce displacements too small relative to localisation uncertainty; long intervals can allow features to deform or leave their comparison regions. There is no universal best separation for every scene. The chosen interval should resolve useful motion while retaining sufficient pattern correspondence, and sensitivity should be examined.
CHAPTER 5 OF 12
5. Remove camera motion without removing river motion
An aircraft or moving mount can shift the entire scene between frames. Apparent image displacement then combines camera motion and water-pattern motion. Stabilisation uses appropriate stationary references to account for the former. A background that looks approximately steady to a viewer can still move enough to bias a small displacement estimate. Quantify residual motion rather than relying only on the impression of a stable video.
In a one-dimensional teaching example, a bank feature shifts three pixels right between frames, while a water pattern shifts fifteen pixels right. Under a simple shared-translation model, relative water displacement is twelve pixels right. The subtraction removes the common image translation. It does not correct rotation, perspective change or depth-dependent parallax, so the simple arithmetic should not be mistaken for a complete stabilisation algorithm.
A useful independent check holds out some stationary reference points from the stabilisation fit. Their residual displacement can reveal whether the corrected scene is actually stationary where it should be. Fitting a transform to a set of points and then reporting perfect agreement at those same points is a weaker test. Choose controls that are genuinely stationary; moving vegetation is not automatically a fixed bank reference.
CHAPTER 6 OF 12
6. Distinguish water-following features from waves and glare
Foam or other naturally occurring surface features may provide useful flow indicators, but their suitability is not guaranteed. Wind can affect surface motion; features can aggregate, dissolve or move differently from the water of interest. A propagating wave crest and a parcel of water do not generally have the same velocity. A method that follows a crest accurately can therefore estimate the wrong physical quantity.
Reflections depend on illumination and viewing geometry. A bright patch can shift because the surface orientation changes, even if it does not represent a material feature moving through the reach. Compare feature behaviour across frames and spatial regions. Where the interpretation is doubtful, mark the affected vectors or areas as uncertain rather than assuming that a high correlation score establishes water-following behaviour.
For transfer, imagine the same measured image displacement on a calm day and a windy day. The numerical conversion to metres per second is unchanged if geometry and timing are unchanged. The confidence that the surface feature represents the required water velocity may change. This separates a measurement calculation from the physical justification for using its result in a discharge estimate.
CHAPTER 7 OF 12
7. Choose comparison regions and reject unsupported vectors
An interrogation region that is too large can average across a shear zone, combining fast main-channel motion with slow marginal water. A very small region may contain too little useful texture for a reliable match. The region size and overlap influence the spatial support of each vector. A dense arrow map does not mean that every arrow represents an independent, equally resolved measurement.
Filtering should use transparent quality criteria and physical context. Outliers can result from poor matches, but unexpected reversals or slow zones can also be real. Automatically deleting every vector that disagrees with a preferred downstream direction can erase recirculation. Preserve the unfiltered output, flags and reasons for exclusion so a later reviewer can distinguish image failures from actual hydraulic complexity.
Try a changed-context check in which the comparison region straddles a quiet pool edge and a fast thread. Shrinking it may reveal separate motions previously merged into one. That does not necessarily mean the river changed between calculations. The observation support changed. Report settings and examine whether the quantities used in the final cross-section calculation remain representative of the intended subareas.
CHAPTER 8 OF 12
8. Use velocity normal to the measuring section
Discharge through a section depends on the velocity component perpendicular to that section, not the full horizontal speed if flow crosses it obliquely. If a surface vector has speed u and makes angle θ with the section normal, its normal component is u cos θ. Define the angle relative to the normal explicitly. Using the angle to the section itself would require a different trigonometric expression.
For a teaching vector of 2.0 m/s at 30 degrees to the normal, the normal component is about 1.732 m/s. Using the full 2.0 would overstate the normal component by about 15.5% relative to 1.732. A channel-aligned camera view does not by itself establish that every local vector is normal to the selected section. Bends and secondary circulation can make direction important.
The section must also correspond to the bathymetry and hydraulic assumptions used in the calculation. A velocity map measured upstream cannot simply be combined with a deeper cross-section downstream without a justified connection. State where the section lies in the image and when its geometry was measured. Spatial alignment is part of the discharge evidence, not merely a plotting convenience.
CHAPTER 9 OF 12
9. Convert surface motion to a depth-averaged estimate cautiously
A simplified discharge calculation sometimes uses a surface-to-depth-mean factor α, giving mean normal velocity approximately α times the observed normal surface velocity. The factor is not a universal constant. Flow structure, roughness, depth and local conditions can affect it. Use site-specific evidence or a justified model, and test sensitivity where the conversion materially affects the result.
For an illustrative area of 6 m² and normal surface velocity 1.5 m/s, α = 0.85 gives 7.65 m³/s. Keeping the observations fixed but using α between 0.75 and 0.95 gives 6.75–8.55 m³/s. This range is a sensitivity scenario, not a recommended factor interval or a statistical confidence interval. It isolates one assumption while holding geometry and image interpretation fixed.
A single factor can also conceal variation across the section. Slow margins, a deep main channel and a shallow bar may need different treatment. More sophisticated conversion methods exist, but their additional structure must be supported by data and valid assumptions. A complex model does not remove the need to explain which velocity is observed and which part of the flow field remains inferred.
CHAPTER 10 OF 12
10. Integrate area and velocity without multiplying incompatible averages
Divide a teaching cross-section into subareas A_i and assign each a representative depth-averaged normal velocity v_i. Approximate discharge by the sum of A_i v_i. This makes the spatial weighting visible. An unweighted mean of velocities multiplied by total area is equivalent only under suitable conditions, such as equal subareas. A velocity map with many samples near one bank can otherwise distort the mean.
Suppose two subareas are 2 and 6 m², with depth-mean velocities 0.5 and 1.5 m/s. Discharge is 2 × 0.5 + 6 × 1.5 = 10 m³/s. The unweighted velocity mean is 1.0 m/s; multiplying it by the 8 m² total gives only 8 m³/s. The missing 2 m³/s is a weighting error, not an image-processing error.
Bathymetry can become outdated through bed movement or changing water level. If area is wrong, a perfect velocity estimate still produces a biased discharge. Keep survey date, reference level and section alignment with the image record. For transfer, raise stage enough to wet a previously dry margin: the new area and its velocity need representation rather than simply reusing the old total area.
CHAPTER 11 OF 12
11. Validate the complete chain with genuinely new evidence
Independent acoustic, current-meter or appropriate tracer measurements can help test image-derived discharge, provided their timing and spatial support are compatible. A difference between methods is not automatically evidence that the image method failed. Each has assumptions and uncertainties. Compare what was observed, what was inferred and which conditions changed between measurements before deciding how to interpret disagreement.
Do not calibrate α to one reference discharge and then use agreement with that same discharge as independent validation. It demonstrates that the chosen factor can reproduce the calibration target. Testing another suitable period or a held-out measurement provides a stronger assessment of transfer. Preserve which observations were used to choose settings and which were reserved to test them.
The USGS research and data releases linked below document image-based surface-velocity work and comparisons with conventional measurements. Their results establish that the approach is a real measurement method, not that any available phone clip is suitable. The examples here deliberately separate image scale, timing, feature interpretation, hydraulic conversion and cross-section integration so that each can be checked on its own.
CHAPTER 12 OF 12
12. Report the result at the level the evidence supports
A useful report may conclude that surface velocity was resolved in the central channel but discharge remained uncertain because the section geometry or depth conversion was weak. That is a legitimate partial measurement. Avoid presenting a complete discharge merely because a software package can export one. Label observed surface vectors, inferred depth-mean values and excluded regions separately.
Keep the original imagery, timestamps, camera and lens information, control geometry, water level, processing settings, vector filters, bathymetry and conversion assumptions. Record units and coordinate directions. A result that depends on an undocumented manual deletion cannot be reproduced reliably. A clear provenance chain lets a later improvement in geometry or filtering revise the estimate without repeating every part of the observation.
Your final transfer test changes the camera angle, the surface texture and the cross-section shape separately. Explain which stage of the chain needs revision in each case. If all three changes are answered only by adjusting α, the model is hiding distinct errors inside one parameter. Sound reasoning repairs the affected measurement stage rather than asking a final correction factor to absorb every mismatch.
Independent worked case and evidence checks
Consider an original fictional example with an approximately rectified overhead view of a small river section. The timing record shows that the frames compared are 0.25 s apart. A stationary bank reference moves two pixels right in the raw images. A central water pattern moves eighteen pixels right. Assume, only for this first calculation, that shared translation is the relevant camera-motion correction.
Corrected image displacement is eighteen minus two, or sixteen pixels right. At a local water-surface scale of 0.020 m per pixel, displacement is 0.320 m. Dividing by 0.25 s gives surface speed 1.28 m/s. Every input has a separate meaning: the two-pixel correction is camera motion, the scale converts image distance, and the time interval converts displacement into a rate.
If camera movement were ignored, eighteen pixels would imply 0.360 m and 1.44 m/s. That is 12.5% above the corrected 1.28 m/s. The error would propagate into a simple linear discharge conversion. Reporting more significant figures would not fix the omitted motion correction. The appropriate repair is at the stabilisation stage.
Now suppose the vector crosses the selected section at 20 degrees to its normal. Normal surface velocity is 1.28 cos 20 degrees, approximately 1.203 m/s. The full speed remains 1.28 m/s, but that is not all directed through the section. Keep both quantities labelled so a future user cannot accidentally use full speed where a normal component is needed.
For an illustrative α of 0.82, the corresponding depth-mean normal estimate is about 0.986 m/s. With a compatible subarea of 3 m², its discharge contribution is about 2.959 m³/s. The factor is an assumed teaching value, not a field recommendation. The arithmetic is transparent precisely so that another justified factor can be substituted without changing the observed image displacement.
Complete a three-part section. The left, central and right subareas are 2, 3 and 1 m². Their normal surface velocities are 0.60, 1.203 and 0.40 m/s. Illustrative conversion factors are 0.80, 0.82 and 0.75. Contributions are 0.96, about 2.959 and 0.30 m³/s, making total discharge approximately 4.219 m³/s.
An unweighted average of the three depth-mean normal velocities is approximately 0.589 m/s. Multiplying by the 6 m² total gives about 3.532 m³/s, not 4.219. The unequal subareas explain the discrepancy. The correct area-weighted mean is approximately 0.703 m/s. A dense image grid should not tempt you into forgetting that discharge is an area integral.
Change the local spatial scale in the central region from 0.020 to 0.022 m per pixel after a geometry review, leaving timing and angles unchanged. Its inferred surface speed and discharge contribution both rise by 10% under the simple linear model. The total rises by about 0.296 m³/s, not by 10% of the whole section, because only the central region’s scale was revised.
Next change the time evidence. If the true interval is 0.20 s rather than 0.25 s, all velocities computed from the same displacements rise by a factor of 1.25. Under unchanged areas and conversion assumptions, total discharge also rises by 25%. This is a shared timing error across the section. It is not reduced by averaging more vectors from those same frames.
For a feature-validity test, imagine the right-hand region tracks a wave crest travelling upstream while visible foam travels downstream. The image calculation could accurately describe the crest and still be inappropriate for a water-following velocity estimate. Flag the region and seek suitable independent evidence. Do not reverse the sign merely to make its contribution agree with the desired total.
For an uncertainty exercise, hold image motion, geometry, timing and direction fixed but vary the central α from 0.77 to 0.87. Its contribution changes by about 0.180 m³/s below or above the central estimate. This isolates conversion-factor sensitivity. It does not include uncertain area, camera stabilisation or feature interpretation, so it should not be labelled the complete uncertainty of the section discharge.
An independent measurement at a compatible time reports 4.4 m³/s with its own stated uncertainty. The difference from 4.219 is about 0.181 m³/s, or roughly 4.1% of 4.4. That comparison can motivate examination of the uncertainty budgets. It does not prove the image method accurate to 4.1% everywhere, and it must not be reused as independent validation if the α values were selected to match it.
Finally, a later flood changes bed shape and covers the old bank control points. Reusing the original processing settings without checking the new geometry would transfer a workflow beyond its evidence. Identify the missing controls, updated section and changed surface conditions. Some parts of the method remain useful, while others require new observations before a comparable discharge can be reported.
Your independent answer should recover sixteen corrected pixels, 1.28 m/s central surface speed, approximately 1.203 m/s normal surface velocity and approximately 4.219 m³/s total discharge under the stated assumptions. Then explain why a precise answer does not remove the need to justify the tracked features, depth conversion and surveyed area. Those distinctions are the real learning target.
Worked comparison
| Stage | Observed or inferred quantity | Main check |
|---|---|---|
| Image comparison | Pixels per frame interval | Useful pattern correspondence |
| Geometry and timing | Metres per second at surface | Water level, scale, timestamps |
| Direction | Normal surface component | Section orientation |
| Hydraulic conversion | Depth-mean normal velocity | Justified conversion method |
| Area integration | Discharge | Compatible bathymetry and weighting |
Questions and practical checks
Can an ordinary video measure discharge?
Only if its geometry, timing, visible flow features and hydraulic conversion are adequately supported. A clip may support surface-motion observations without supporting a complete discharge estimate.
Does a high correlation prove that a feature follows water?
No. It supports an image match. Waves, reflections or repeating patterns may still represent the wrong physical motion.
Why is bathymetry needed?
Discharge requires flow over a wetted cross-section. Surface velocity alone does not provide the area or the vertical velocity distribution.
Is the surface conversion factor always 0.85?
No. Any factor needs justification for the site and conditions. The values in this guide are illustrative assumptions, not universal defaults.
What is the strongest quick check?
Recompute one region from displacement, elapsed time, scale and direction, then trace how its area-weighted contribution enters the section total.
Continue with the connected subjects
Compare independent discharge evidence
Moving-boat acoustic discharge and navigation-reference errors
Index velocity, section area and reversing flow
Tracer-dilution discharge and independent mixing assumptions
Sources and scope
USGS research: discharge estimation using video recordings from small unoccupied aircraft systems
USGS data release: image velocimetry and Doppler radar measurements for river velocity and discharge
The numerical cases in this guide are original fictional teaching examples, not measurements from a real site. This is an educational explanation, not a field safety instruction, engineering approval or an official certification. For a real project, use qualified practitioners, current product documentation and site-specific evidence. Sources were checked on 30 October 2026.