Improving Accuracy and Robustness of Space-Time Image Velocimetry (STIV) with Deep Learning

Round 1

Reviewer 1 Report

The topic of the present paper is interesting under several viewpoints: understanding the overall river dynamics, design of hydraulic structures, preventing urban flooding, etc. The use of deep learning to improve the capability of the STIV software is a good idea, and the illustrated examples are suitable for solving the main weaknesses of the software itself.

The paper is well organized, but some points should be better addressed, as illustrated in the following.

 

All along the text, it should be better to homogenize symbols and units: specifically, the symbol “ ° ” is suggested in place of “deg” or “degrees” (e.g., Fig.3, L194, L195, L211, L212)

L54: it should be “time- and line-averaged”.

L60: it should be “weather conditions AND during …”.

L80: remove dot after “Figure 2”.

L104: please, clarify the practical meaning of “channels”, “height” and “width” in relationship with the gradient \phi.

L130-131: this statement should have a robust basis, hence a suitable reference is needed here.

Fig.3: instead of “Angle”, the appropriate symbol must be used (\phi or \tilde{\phi}).

L135-136: this sentence could be improved as “Next, a 2D Fourier transform IS PERFORMED and a high-pass filter IS APPLIED to remove …”.

L211: “ANGULAR analysis”.

L216-218: for the sake of clarity, the Gradient Tensor method should be briefly presented here or in a separated subsection of section 2.

Section 3.3: the organization of this section should be completely revised, as there are many unclear or missing points:

• there is a jump between L227 and L238, this leading to an awkward reading of the paragraph;
• the sentence at L238-240 is unclear and should be reworded;
• in fig.8, the blue lines cannot be easily detected;
• figures 8-11 should be better arranged, e.g. moving them some lines after L244 (where they are recalled); on the other hand, Tab.4 should be placed between L243 and L244 (in any case, before figures 8-11);
• all locations illustrated in figures 8-11 should be briefly described in the text and also reported in the captions;
• in contrast to the comment at L248-249, I can see that most of the oscillations/instabilities exist close to the left bank (distance < 30m), am I right?
• the comments related to the other methods illustrated in fig.12 should be improved by: 1) briefly describing the used methods (here or in a dedicated subsection of section 2), 2) providing some statistical analysis which highlights the goodness/badness of all used methods.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

GENERAL COMMENTS

This manuscript discusses the potentials of deep learning to improve the accuracy and robustness of the space-time image velocimetry (STIV) technique. More specifically, the authors suggest using a convolutional neural network (CNN) in order to estimate the angle of the inclination of the patterns in the space-time images (generated by stacking the pixels of a search line along the time axis). The idea is interesting and useful, since it appears to improve the efficiency of the STIV method. However, the manuscript requires improvements:

- Some sentences are vague and require clarification (see specific comments).

- Figure 12 demonstrates that, under adversary conditions, there are significant differences among the evaluated methods. There are some recent studies concerning the uncertainty introduced in the image velocimetry methods due to adversary conditions, which should be taken into account by the authors. For example Deter (2021) has extensively studied various sources of uncertainty and errors in the image velocimetry methods (e.g., misleading small vectors because of either smooth water surfaces or moving objects like shrubs on a windy day). Rozos et al. (2020) have employed Monte Carlo simulations to analyze the influence of the uncertainty of the parameters that control the sensitivity of the LSPIV on the efficiency of this method. With these studies as background, the authors need to add some discussion about the results appearing in Figure 12.

- Some restructuring is required. The lines 117-131 should go to sub-section 2.1. The sub-sections 2.2, 2.3 (without the lines 117-131) and 3.1 should be merged into a single section (or sub-section) placed after 2.1. This section should provide all the relevant information about the topology of the employed CNN and how it was trained/verified.

- The 2D Fourier transformation is not explained clearly. In fact, the text creates confusion whether this is a novel concept of this study or a previously introduced and tested technique (see specific comments).

- The authors should discuss what are the CPU requirements of their method (CNN) compared with the alternative approaches (gradient tensor method, QESTA, etc.) and what is the overall time required to run STIV.

 

 

SPECIFIC COMMENTS

Location: "As a result, it became clear that the STIV method using deep learning often gives correct results when the conventional method outputs outliers."
Comment: This wording ("often gives correct results") raises many questions. Maybe it would be better to write "The case studies indicated that deep learning can improve the efficiency of the STIV method."

Location: "...flow measurement methods such as the ADCP (Acoustic Doppler Current Profiler) [4-6], radio wave anemometers [7,8]..."
Comment: An ‘anemometer’ is a device that measures the wind speed. How can wind speed be related to this study?

Location: "The advantages of STIV over other methods are that it can analyze images even with a small depression angle of about two degrees by paying attention to time-and line averaged streamwise velocity component."
Comment: This is not clear. STIV presupposes knowledge of the direction of the flow. This is not straightforward in every case study, so definitely this is a disadvantage of STIV.

Location: "f(I) , which calculates the gradient φ of the stripe pattern from the image intensity value information I ∈ R ch×H×W of the STI (channels: ch , height: H , width: W )"
Comment: This is not clear. The height and width are supposed to be the number of frames and the length of the search line? Please define more clearly the dimensions of I.

Location: "... and the numerical simulation by DNS [32], ..."
Comment: The abbreviation "DNS" has not been introduced.

Location: "Also, we used the two-dimensional Fourier transform image [30] of the STI [23], rather than the original STI, as the input to the CNN. ... Using various types of STIs, we have confirmed that using the Fourier transform image of the STI as an input has better gradient detection performance than using the original STI."
Comment: Has the performance of this Fourier transformation been introduced and evaluated by other researchers or this concept is tested in image velocimetry for the first time here?

Location: Figure 2 and the text in lines 119-131
Comment: The authors use the term 'turbulence' to describe the signal corresponding to the flow. This is confusing both because turbulence is a source of noise (and uncertainty) in some cases and because artificial seeding or other types of natural features (e.g., debris, boils,etc.) are the targeted characteristics in other cases. It would be more clear to use the term ‘flow signal’ instead.

Location: Table 4
Comment: What is Δt?

Location: "In this study, we applied a deep learning method, which has greatly developed in the field of image analysis, ..."
Comment: What is meant with 'greatly'?

 

 

REFERENCES

Detert, M. (2021). How to avoid and correct biased riverine surface image velocimetry. Water Resources Research, 57, e2020WR027833.
https://doi.org/10.1029/2020WR027833

Rozos, E.; Dimitriadis, P.; Mazi, K.; Lykoudis, S.; Koussis, A. On the Uncertainty of the Image Velocimetry Method Parameters (2020). Hydrology, 7, 65. https://doi.org/10.3390/hydrology7030065

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The STIV has great potential for real-time monitoring of rivers due to its high spatial resolution and low time complexity. Still, the generated space-time image includes many noise and interference textures, making practical applications inevitable. This study produces meaningful results that contribute to the accuracy of STIV using CNN. To accept the paper, the authors need to make some of the following corrections:

- Avoid using abbreviations in titles. Use the full name of “STIV.”

- Line10-22, Reduce descriptions in Abstract and include key results.

- Line26-32, Add relevant references to the first and second sentences of the Introduction.

- Line 31, add a comma before “even.”

- Line 32, missing a comma after “in Japan,” modify “a float” to “floats.”

- In section 2.1, the phrases used in Fujita et al. (2020) were used as-is without quotation.

- Add river flow direction information in Figure 1. 

- Line 80, “an example of the original STI,” … There are many cases of unnecessary use or missing articles throughout the paper.

- For Figure4, provide a higher resolution figure.

- The numerical information was cut out in Figure 14. 

- Add a discussion section, and conclusions should be reinforced as a whole. 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors have carefully addressed all my comments.