the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 07 Jan 2025
Unsupervised image velocimetry for automated computation of river flow velocities
Abstract. Accurate, long-term, measurements of river flow are imperative for understanding and predicting a broad range of fluvial processes. Modern technological advances are enabling the development of new solutions that are tailored to manage water resources and hazards in a variety of flow regimes. This study appraises the potential of freely available image velocimetry software (KLT-IV) to provide automatic determination of river surface velocity in an unsupervised workflow. In this research, over 11,000 videos are analysed, and these are compared with 1-D velocities derived from 303 flow gauging measurements obtained using standard operating procedures. This analysis was undertaken at a complex monitoring site with a partial view of the channel with river flows spanning nearly two orders of magnitude. Following image velocimetry analysis, two differing approaches are adopted to produce outputs that are representative of the depth-averaged and cross-section averaged flow velocities. These approaches include the utilisation of theoretical flow field distributions to extrapolate beyond the field of view, and an index-velocity approach to relate the image-based velocities to a section averaged (1-D) velocity. Analysis of the section-averaged velocities obtained using KLT-IV, compared to traditional flow gauging, yields highly significant linear relationships (r2 = 0.95–0.97). Similarly, the index-velocity approach enables KLT-IV surface velocities to be precisely related to the section-averaged velocity measurements (r2 = 0.98). These data are subsequently used to estimate river flow discharge. When compared to reference flow gauging data, r2 values of 0.98 to 0.99 are obtained (for a linear model with intercept of 0 and slope of 1). KLT-IV offers an attractive approach for conducting unsupervised flow velocity measurements in an operational environment where autonomy is of paramount importance.
- Preprint
(1431 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 18 Feb 2025)
-
RC1: 'Comment on hess-2024-213', Anonymous Referee #1, 19 Jan 2025
reply
The manuscript deals with image processing aimed at obtaining reliable flow discharge estimates under continuous unsupervised operation. The utility of this approach is acknowledged by the WMO, recently issuing an update to WMO-No. 168, section 5.3.7.3 – Image-based methods for discharge measurements. The authors add to the general discussion on the applicability, accuracy and associated uncertainties by presenting the results from nearly one-year continuous observations on the River Dart, UK. The data were collected at Austins Bridge, where available hydrometric records span back to 1958, and treated in compliance with the WMO-No. 168 requirements.
The River Dart at Austins Bridge yields the longest hydrometric record on the River Dart, but its hydrologic presentation in subsection 2.1 is apparently scarce. It seems however important to put the experimentation in the proper context, in terms of both average and extreme flows. The notion on ‘reference observations’ appears as early as in Subsection 2.5.1, where they are briefly mentioned in L172 as a source of the conversion factor, but the details are lacking. What is their distribution across the velocity/discharge range? What are the standard data collection protocols? Even the number of observations is uncertain. The total number of reference observations (303) is presented in the Abstract, but never discussed elsewhere. Were some of the, or all aDcp observations (at least some were, as stated in L182, but clearly not all, since the record starts is 1981)? Subsection 2.5.1 is vague on the subject, L177 has no notions on the Ua protocol, and in the rest of the subsection only aDcp data are discussed.
The manuscript is generally well-written, and most issues that can be reported are related to an overall quality of presentation and text flow. While the manuscript grosso modo has most of the data needed, some of these data are scattered across the manuscript making it hard to follow. Several by-line comments are also given below:
L31: are these mentioned below algorithms, in the sense of Kanade-Lukas algorithm, or rather methods and approaches, as per WMO-No. 168?
L41: reference the Lukas-Tomasi algorithm, and probably mention Kanade here also, since KLT in KLT-IV might presumably stand for ‘Kanade-Lukas-Tomasi’ (indeed, references from L132 can be moved up here).
L43: PIV is not presented above, but is apparently the same as “PIV” in LSPIV, for the latter is the same PIV in the narrower context (WMO-No.168).
L43-44: in referencing (Tauro et al., 2018b) it can be noted that at least some of the results of that paper were obtained using “woodchips … continuously deployed” from the upstream so not fully in the context of a “continuous, automated, and unsupervised workflow”.
L94-95: this phrase evokes the question on what were the hydrological conditions during the experiment, of which there is no notion in Section 2.1. How they are compared to long-term averages and extremes? It is only later in the manuscript (L178) that we will know that the reference flow measurements range from 1.7 to 145 m3s-1; no info is provided on the flow discharge ranges during the experiment.
L170: “20 segments of equal size”; from Figure 2, these are segments of equal width, not equal area, though for polygons the latter is most commonly assumed. This needs to be stated explicitly. Am I right to understand this so that while the flow width increases with the increasing discharge, the width of each segment increases accordingly?
Also, this first paragraph of the Subsection 2.5.1 is somewhat confusing. First, while the basic references are provided, I would be interested in having more details on how exactly the observed data were extrapolated over the regions out of the camera sight using these three techniques; why there are 20 segments, and not more/less – how is this justified? In L170, was it so that “average velocities for 20 segments of equal sizes are established” stands for the water surface velocities averaged across the segment width? What is the number of pixels with data in each segment surface? How exactly was alpha derived? The (Perks, 2024b) reference is misguiding in that might presumably lead to the supportive information while it simply references a figure derived from unexposed data. Were the alpha values derived segment-wise for all 60 aDcp measurements? Were the segments similar or close in size (width?) in aDcp and extrapolation exercises? Further on, is the “cell” in L173 equivalent to “segment” in the lines above? If not, the difference must be explained otherwise the text is ambiguous.
L180: In “calculate the (1-D) velocities derived from reference observations”, can these be replaced simply by Ua, as given in L177? Similar to this, can “1-D” be omitted from most instances where its presence is misleading or redundant, e.g., L184, L187, L191-192, L198 and almost each and every instance below. Once the definition is given, there is no need to reiterate, and it makes sense to substitute. As in L180, same in L196-197, same in L198-199, and so on.
In this Subsection 2.5.2, again, the statistics for reference stage observation can be highly utile. Were these reference observations mentioned in L196 the same as presented elsewhere including the Abstract? Can the authors consider providing the details either here or in the Section 3?
L210: It is probably more correct to not start the “Velocity reconstruction” section with the overview of the EA data. Also, in Figure 3, velocity reconstruction is applicable to both procedures employed, see Subsections 2.5.1 and 2.5.2 accordingly. Thus Section 3.1 must be renamed accordingly.
L220: more justification is needed here or above on claiming the constant Froude as a best-performing model, since the Figure 4 is indecisive on this matter.
Figure 5: is it correct that the distances are given from the left bank while the camera is on the right bank?
L244-245: how exactly the conclusion that “as few as eight flow gauging measurements” suffice to successfully quantify the cross-section averaged velocity was reached? Does it imply that the uncertainty (variability) level is acceptable at this n.
L266: “autonomously analysed in an unsupervised workflow”. This said, it must be noted however that while the imagery was indeed collected in that way, its further treatment and the final output weren’t possible without the prior work on-site, including aDcp surveys under challenging conditions of the below 1% flood. Without this, based solely on the a priori knowledge of the site, what the potential errors of the standalone application could have been?
L278-279: Top and bottom blind zones are present in aDcp measurements, along with right and left margins, so the reported Q is corrected for these in RiverSurveyor software. The wetted cross-section area has some effects on the aDcp-derived flows, but it is not to be overestimated.
Overall, I concur with the authors in that the major utility of the described approach is in the “development or refinement of stage-discharge rating curves”, where the OTV and OTV-like approaches can be highly utile in increasing the accuracy of high flow estimations. For the River Dart at Austins Bridge, the highest observed peak flow is 550 m3s-1 (1979-1980), almost 400 m3s-1 over the observed highest peak flow. At this part of the flow duration curve, as far as I understand, the OTV can substantially improve the flow estimates in the straight single-thread non-deformable channel.
Citation: https://doi.org/10.5194/hess-2024-213-RC1
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 2,401 | 24 | 1 | 2,426 | 0 | 0 |
- HTML: 2,401
- PDF: 24
- XML: 1
- Total: 2,426
- BibTeX: 0
- EndNote: 0
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1