Crack Detection and Displacement Measurement of Earth-Fill Dams Based on Computer Vision and Deep Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The research point is good and it raises important points in the efficiency and durability of earth dams exposed to a force that causes cracks in the dam body. During the work, the test results were verified and some available models of the width of cracks in the body of the earthen dam were compared. The research may be accepted, but the following questions must first be answered:

• Language review on line 257.
• In Figure 7, it is clear that the dam was flooded with water and that the water drained from the top of the dam to the bottom, creating a hydraulic jump. Here come two questions: To what extent does flooding the dam affect the appearance and size of cracks in the dam body? Secondly, what is the extent of the impact of the appearance of the hydraulic jump on the appearance of cracks and their size?
• It is well known that flooding dams with water will affect their efficiency, especially since they are made of loose earth. Were cracks measured before or during flooding? And what is the difference between the two values?

Author Response

Thank you for the valuable comment and constructive suggestion. Our responses are included in the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper mainly studies the crack detection and displacement measurement of earth dams based on computer vision and deep learning technology. The author proposes an improved YOLOv8-CGA crack segmentation method, which introduces the cascaded group attention mechanism (CGA) to enhance the accuracy of crack detection. Additionally, the article presents a width calculation method based on skeleton lines to improve the accuracy of crack width measurement. Dams are very common in various countries and regions and have significant agricultural, engineering, and ecological significance. The detection and repair of different types of dams, especially earth dams and rock dams, are of great importance. This paper has a clear research purpose, but before it can be further processed, some questions need to be answered and solved. My main concerns are as follows:

1.Although some existing methods are time-consuming and labor-intensive in implementation, methods such as ultrasonic can determine internal hazards and damages of structures, especially in some dams where the interior is completely damaged but there are no cracks or only minor cracks on the surface. This situation also exists objectively. How does the author consider this issue? I am more concerned that the author should further analyze the relationship between the extraction results of surface cracks and the internal structure to provide more authoritative conclusions.

2. In Section 2.1, the description of the crack feature parameter extraction process is not detailed enough, and there is no clear theoretical and methodological support. There is also repetition with the subsequent parts, which should be improved.

3. The crack extraction scheme in the paper is based on image processing and deep learning. The existing image processing methods for extracting geometric or texture features of cracks do not seem very complex. For example, there are a series of image and graphics processing algorithms and skeleton extraction schemes in MATLAB software, and they have obvious accuracy. So where is the author's innovation? Has a comparison been made with traditional schemes? I suggest adding a comparative study.

4. The author should specifically provide the types of noise and processing schemes in the crack extraction process of different types of dams. How are the crack extraction results interrupted due to factors such as shooting conditions, physical obstructions, and resolution solved? Is it just simple dilation operation? In this case, skeletonization processing may cause the generation of burrs and changes in single-pixel skeleton paths. How does the author consider and solve these problems?

5. It is not recommended to use "crack width" as a keyword.

Author Response

Thank you for the valuable comment and constructive suggestion. Our responses are included in the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper presents a comprehensive computer-vision approach for monitoring earth-fill dams, combining YOLOv8-CGA for crack detection with image processing and DIC-based displacement analysis. The overall research logic - progressing from identification, to localization, and finally to estimation - is appropriate, but several important points should be clarified.

1. Crack detection flow: identification ->> localization ->> estimation

Overall, the research structure is easy to follow. Cracks are firstly identified using the YOLOv8-CGA model, then localized through image-processing steps, and finally quantified through skeleton-based width estimation and DIC analysis. They should be consecutive steps instead of parallel steps as shown in the Figure 1.  I would suggestto place the "Displacement measuremet" as a following steo after "Crack detection" to streamline your studies. 

2. The scope and limitations of YOLO

YOLO is a fast, real-time object detection framework that can be applied for objects identification and tracking, but is not designed for feature segmentation or displacement measurement. The authors should clarify why YOLO was chosen instead of networks such as other image processing methods for pixel-level crack analysis.

More importantly, the paper does not explain how the training dataset was prepared such as scale, diversity, and labeling process.

Furthermore, model robustness should be tested under variations of lighting and surface texture conditions. The assessment of the general applicability of the proposed YOLOv8-CGA approach should be explored.

3. Supervised or Unsupervised Learning

From the description, the model appears to have been trained under supervised learning, since YOLO depends on annotated bounding boxes or masks. This should be stated clearly where cracking does not have a clear feature to detect especially at the early stage. If transfer learning, pre-training, or semi-supervised data augmentation was used, the authors should include those details as well. Such information is essential to evaluate the repeatibiluty and generalizability of the proposed method. 

4. DIC for Identification and Prognosis

As stated in thie paper, DIC is primarily used for displacement measurement, but it can do more than that. When DIC images are captured continuously, they can reveal full-field displacement and strain patterns, which are extremely useful for detecting crack initiation, tracking propagation, and even forecasting failure, so called prognosis. Within this regime, the paper would be stronger if this broader potential of DIC can be acknowledged. In early stage prognosis, visualizing deformation or gradient of deformation trends is often more valuable than focusing purely on stress or strain calculations.

5. Limitations of YOLOv8-CGA

The study should include the discussion of the model’s limitations. For example, 

• Scale sensitivity: very fine or micro-cracks may not be detected reliably.
• Environmental dependence: factors like lighting conditions, types of soil or surface textures, etc. 
• Dataset generalization: results from controlled laboratory settings may not be transferrable directly to real world conditions.

Addressing these points would be very helpful to understand the strengths and constraints of your proposed method.

6. Literature Coverage

Most of the cited reference are very recent studies and many foundational key contributions in YOLO, SHM, and DIC research are not included. Classic studies on YOLO, DIC fundamentals, feature-based crack detection and structural health monitoring (SHM) should be cited. Including these studies will show the continuity in the research field and demonstrate that the study was built upon established knowledge rather than only on current trends.

7. Overall Evaluation

The paper presents an interesting view to use YOLO for crack detection and monitoring on earth-fill dams and sustainable infrastructure. However, the manuscript can be furture improved with a clearer explanation of its workflow, stronger justification the selection of YOLO, providing more details of training and sensitivity analysis, and including more classic referencing. 

 

Author Response

Thank you for the valuable comment and constructive suggestion. Our responses are included in the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

After all comments have been addressed, the paper can now be accepted for publication.