Review Reports - Seed 3D Phenotyping Across Multiple Crops Using 3D Gaussian Splatting

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Comments to the authors The 3DGS method produces denser point clouds and higher-quality reconstructions, yet it required 2.3× the reconstruction time of SFM-MVS at 30,000 iterations. This could be a problem for scalability how are you going to overcome it/

The study includes multiple crop types, however the number of seeds or replicates tested per variety is not specified so please provide the number since this is critical for statistical robustness. based on this remark the authors do not present statistical tests (e.g., t-tests, ANOVA) to determine whether improvements in R², MAPE, or RMSE between methods are statistically significant.

The authors also should use more differnet seeds

Author Response

RESPONSE: We greatly appreciate your meticulous review of our manuscript and the valuable suggestions you have provided! Your feedback has been crucial in enhancing the clarity and reproducibility of the paper. Based on your recommendations, we have made comprehensive revisions and added further explanations.

We sincerely thank you for each of the constructive comments you have provided, which have significantly contributed to improving the rigor and scientific quality of our manuscript. We hope that the revised manuscript now presents the research content more clearly and completely, and we look forward to your further review and valuable feedback!

We will respond to and improve the following aspects one by one:

Comments to the authors The 3DGS method produces denser point clouds and higher-quality reconstructions, yet it required 2.3× the reconstruction time of SFM-MVS at 30,000 iterations. This could be a problem for scalability how are you going to overcome it/

RESPONSE: Thank you for your advice! Thank you for your insightful comment. We agree that reconstruction efficiency is an important consideration for scalability. Although 3DGS required approximately 2.3× the reconstruction time of SFM-MVS at 30,000 iterations, our experiments show that comparable reconstruction quality (PSNR > 34 dB, SSIM > 0.93) can already be achieved at 7,000–10,000 iterations, effectively reducing the processing time to a level similar to SFM-MVS. This revision is located at line 582 in the attached file.

The study includes multiple crop types, however the number of seeds or replicates tested per variety is not specified so please provide the number since this is critical for statistical robustness. based on this remark the authors do not present statistical tests (e.g., t-tests, ANOVA) to determine whether improvements in R², MAPE, or RMSE between methods are statistically significant.

RESPONSE: Thank you for your valuable suggestion. We have now clarified the number of seeds and replicates for each variety and performed statistical analyses to ensure robustness. Specifically, the seeds of each cultivar were divided into five groups of ten seeds each for independent experiments, resulting in a total of 300 seeds and 30 experimental runs across all cultivars (Line 156).

In addition, paired t-tests were conducted to evaluate the statistical significance of performance differences between reconstruction methods. The results show that the mean RMSE values were 0.2223 for 3DGS and 0.3450 for SfM–MVS (t = –2.539, p = 0.0143), and the mean MAPE values were 1.91% and 2.98%, respectively (t = –2.569, p = 0.0133). These results indicate that the accuracy improvement of 3DGS over SfM–MVS is statistically significant (p < 0.05). (Line 489)

As the coefficient of determination (R²) is computed over the entire dataset rather than per individual seed, it was not included in the t-test analysis.

The authors also should use more differnet seeds

RESPONSE: Thank you for your advice! In the current study, we intentionally selected seeds of maize, wheat, and rice for the following reasons:

Agricultural Significance: As the three major staple foods globally, research on their seed phenotypes is of paramount importance.

Morphological Representativeness: These three types of seeds exhibit high diversity in size (from large to small), shape (from near-spherical to elongated ellipsoids), and surface texture (from smooth to grooved), effectively representing a wide range of phenotypic characteristics common in cereal grains.

We would like to emphasize that the phenotyping pipeline proposed in this paper, based on 3D Gaussian Splatting, is a general framework. Its core strength lies in its powerful capability for reconstructing object geometry and appearance, rather than being tailored to a specific seed type. Our current experimental results have demonstrated that this technique can simultaneously and accurately handle these three seed types with significant differences in size, shape, and texture, which in itself provides preliminary evidence of its good generalizability.

Should the reviewer consider it crucial for the current revision, we would be happy to incorporate a preliminary analysis of soybean seeds to provide immediate evidence of the method's performance beyond cereals.

Once again, we are grateful for your insightful comments, which have significantly helped us improve this work.

Reviewer 2 Report

Comments and Suggestions for Authors

Review report for the article

Agriculture-3910605

The manuscript titled “Extraction of Seed Phenotypes Based on 3D Gaussian Splatting” presents a scientifically relevant and noteworthy investigation which introduces a versatile seed 3D reconstruction method applicable to multiple 10 crops—including maize, wheat, and rice—designed to overcome the inefficiency and subjectivity of manual measurements and the high costs of laser-based phenotyping. These results confirm the method’s effectiveness in plant phenotypic trait quantification and highlight its potential for broader applications, including phenotypic analysis of crops such as wheat and rice. Overall, the manuscript is well-structured and supported by appropriate statistical analyses; however, further elaboration on certain methodological and interpretive aspects could improve the reproducibility and scientific contribution of the work.

Since the study focuses on Extraction of Seed Phenotypes Based on 3D Gaussian Splatting, it would be more appropriate to rewrite the title.
In abstract, Line no 14, what is this SFM? Authors need to put full name before to use any abbreviation. Also give outcome of this study in the abstract.
In introduction, Define the research gap more sharply, and specify what is novel about this study.
Explicit hypothesis and objectives: make a concrete hypothesis.
Literature integration: The introduction could better synthesize recent findings (2020–2024). Authors need to improve it with new references.
How many replications authors take for the analysis? Pls mention.
Improve the figure 7.
Improve the figure 8.
IN figure 10, pls mention the axes and any values.
Authors need to improve discussion part.
Replication and sample size: Biological replication may be sufficient for robust statistical inference.
Ensure that all abbreviations are defined upon first use in the text.
Please ensure that journal names and scientific names in the reference list are properly italicized, following the journal’s formatting guidelines.
There are several instances of inappropriate or unnatural English expressions throughout the manuscript.
Please consider submitting the manuscript for professional English editing or having it reviewed by a native English speaker.

Comments on the Quality of English Language

There are several instances of inappropriate or unnatural English expressions throughout the manuscript.
Please consider submitting the manuscript for professional English editing or having it reviewed by a native English speaker.

Author Response

We will respond to and improve the following aspects one by one:

Since the study focuses on Extraction of Seed Phenotypes Based on 3D Gaussian Splatting, it would be more appropriate to rewrite the title.

RESPONSE: Thank you for your constructive suggestion. We agree that the previous title could be refined to better reflect the study’s focus on high-accuracy phenotypic reconstruction using 3D Gaussian Splatting. Accordingly, we have revised the title to:

“Seed 3D Phenotyping Across Multiple Crops Using 3D Gaussian Splatting”

This new title more clearly highlights both the methodological contribution and the application to seed phenotyping across multiple crops.

In abstract, Line no 14, what is this SFM? Authors need to put full name before to use any abbreviation. Also give outcome of this study in the abstract.

RESPONSE: Thank you for pointing this out. We have revised the abstract to define the abbreviation at its first occurrence as “Structure from Motion (SfM)” and have added the main quantitative outcomes of the study, including the coefficients of determination (R² = 0.9361, 0.8889, and 0.946 for seed length, width, and height, respectively) and PSNR values ranging from 35 to 37 dB. These modifications clarify both the terminology and the key findings. The revised abstract can be found in Line 17 of the revised manuscript. Regarding the study outcome: We have added a concluding sentence to explicitly state the key outcome of our study. The added sentence is: "The key outcomes of this study confirm that the 3DGS-based pipeline provides a highly accurate, efficient, and scalable solution for digital phenotyping, establishing a robust foundation for its application across diverse crop species."( Line 28)

In introduction, Define the research gap more sharply, and specify what is novel about this study.

Explicit hypothesis and objectives: make a concrete hypothesis.

Literature integration: The introduction could better synthesize recent findings (2020–2024). Authors need to improve it with new references.

RESPONSE: Thank you for your advice!

Sharper Research Gap and Novelty: We have now explicitly defined the research gap by stating that *"the applicability and performance of 3DGS for high-precision, seed-scale 3D reconstruction and phenotypic quantification remain largely unexplored"*. The novelty of our study is highlighted as the first to systematically apply and validate 3DGS for this specific purpose across multiple crop species, in direct comparison with SFM-MVS.

Explicit Hypothesis and Objectives: We have formulated a concrete, testable hypothesis: *"we hypothesize that 3D Gaussian Splatting can achieve superior accuracy and efficiency in reconstructing seed morphology and extracting quantitative phenotypic traits compared to traditional SFM-MVS methods."* Furthermore, we have listed three specific, measurable objectives to test this hypothesis.( Line 133)

Thank you for your valuable suggestion. We have expanded the Introduction to incorporate recent advancements in 3D Gaussian Splatting. Specifically, we added GaussianPro (Wu et al., 2024), which improves optimization efficiency via progressive propagation; Mani-GS (Chen et al., 2024), which integrates triangular meshes for geometry manipulation; and PyGS (Li et al., 2024), which employs a pyramidal Gaussian hierarchy for large-scale scene representation. These updates provide a more comprehensive overview of recent progress and emphasize the growing adaptability of 3DGS. (Line 122)

4.How many replications authors take for the analysis? Pls mention.

RESPONSE: We thank the reviewer for raising this important point regarding experimental replication. We apologize for any lack of clarity in the original manuscript.

To provide more detail, our experimental design was structured as follows:

We tested two cultivars each of maize, wheat, and rice.

For each cultivar, 50 seeds were divided into 5 groups of 10 seeds.

Each group of 10 seeds was processed and analyzed independently, constituting one experimental replication.

We have now revised the Materials and Methods section (line 156) to explicitly state the number of replications to avoid any ambiguity. The added text reads:

"To guarantee reliable and reproducible outcomes, the seeds of each cultivar were divided into 5 groups of 10 seeds each for independent experiments. This design resulted in a total of 300 seeds and 30 experimental runs across all cultivars."

Thank you for prompting us to clarify this crucial aspect of our experimental procedure.

5.Improve the figure 7.

Improve the figure 8.

IN figure 10, pls mention the axes and any values.

RESPONSE: Thank you for your helpful comments. We have improved Figure 7 and Figure 8 by refining the axis labels, adjusting the font size, and enhancing the overall clarity of numerical values to make the figures more readable.

In addition, the labeling error of Figure 10 has been corrected to Figure 11, and the figure position has been adjusted downward for proper alignment within the text.

6.Authors need to improve discussion part.

Replication and sample size: Biological replication may be sufficient for robust statistical inference.

RESPONSE: Thank you for your valuable suggestion. The Discussion section has been revised and expanded to provide a more comprehensive interpretation of the experimental results and their implications. Specifically, we elaborated on the comparative performance between 3DGS and SfM–MVS, discussed the potential reasons for performance differences across crop types, and related our findings to previous studies.

Furthermore, we clarified the experimental design to emphasize biological replication and statistical robustness. Each crop cultivar included five independent groups, with ten seeds per group, resulting in a total of 300 seeds and 30 experimental runs across all cultivars. (line 572)

7.Ensure that all abbreviations are defined upon first use in the text.

Please ensure that journal names and scientific names in the reference list are properly italicized, following the journal’s formatting guidelines.

There are several instances of inappropriate or unnatural English expressions throughout the manuscript.

Please consider submitting the manuscript for professional English editing or having it reviewed by a native English speaker.

RESPONSE: Thank you very much for your careful review and constructive suggestions. We have thoroughly revised the manuscript to address the issues mentioned. All abbreviations (e.g., SfM, MVS, 3DGS, RMSE, MAPE, etc.) are now clearly defined upon their first appearance in the text.

The reference list has been completely reformatted according to the official MDPI reference style (MDPI.ens) downloaded from https://www.mdpi.com/authors/references.

In addition, the entire manuscript has undergone professional English language editing through the MDPI Author Services to improve grammatical accuracy, clarity, and overall readability.

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript ID: agriculture-3910605 entitled ”Extraction of Seed Phenotypes Based on 3D Gaussian Splatting’’ by Gao et al. investigated the potential and advantages of applying the 3D Gaussian Splatting (3DGS) method for three-dimensional reconstruction and phenotypic measurement of crop seeds. The idea and intention of the paper seems to be good. However, the paper needs some clarifications, and improvement before acceptance for publication. Below you can find to be considered the relevant points:

Major points

The authors employed two methods SfM-MVS and 3DGS, for constructing high-accuracy 3D models of seed entities. Why the authors didn’t test the others methods mentioned such as LiDAR or NeRF.
Line 104-106, The authors should more explain this sentence ‘’Compared with SfM-MVS, 3DGS also requires an initial sparse reconstruction via SfM; however, its key distinction lies in the sub- sequent dense representation stage’’.
Line 159-161: The conversion formula to HSV is not clear. Any references!
In the data acquisition phase, how the authors extract randomly key frames from the video in the first step?
What is the meaning of ‘’stochastic gradient descent’’?
Equation 8: authors should explain the meaning of ‘’T’’.
The authors should explain how they choose the calibration board?
The authors should more explain how ‘’Alpha Shape’’ is analogous to the ‘’2D rolling ball method’’.
Extensive English language editing is needed.

Minor points

Some abbreviations in the abstract such as SFM; 3DG should be defined.
Line 76: Reference ‘’Zermas et al.’’ should be numbered. The same point in line 95 (Mildenhall et al.
Species should be written in italic. Please check the whole manuscript.
Please refer to the journal instructions when you written the different formula.
The quality of figure 3 should be improved.
The legend of figures 4 and 5 should be more detailed.
Please check the legend of figure 6.
The figure 10 should be placed after the paragraph of line 484-488.
Many punctuations and typos errors. Please check the whole manuscript.
References list should be checked and revised accordingly to the journal instructions (see references 28; 30; 32; 33).

Comments on the Quality of English Language

Extensive English editing is needed.

Author Response

We will respond to and improve the following aspects one by one:

The authors employed two methods SfM-MVS and 3DGS, for constructing high-accuracy 3D models of seed entities. Why the authors didn’t test the others methods mentioned such as LiDAR or NeRF.

RESPONSE: Thank you for your insightful comment. Our study primarily focuses on developing a low-cost and easily deployable 3D seed reconstruction pipeline using consumer-grade mobile devices. Therefore, LiDAR-based reconstruction was not included in this work, as LiDAR systems require specialized hardware and higher acquisition costs, which contradict the practical goal of providing a simple, scalable, and accessible phenotyping solution. Instead, we used video acquisition from smartphones to evaluate the reconstruction accuracy of 3D Gaussian Splatting (3DGS) against manual ground-truth measurements, which is sufficient to validate the effectiveness of the proposed approach under low-cost imaging conditions.

Regarding NeRF, we note that Instant-NGP, which is included in our experiments (see Table 5), is a highly efficient implementation of NeRF that incorporates hash-encoding and GPU-accelerated training. Hence, the comparison between Instant-NGP and 3DGS effectively reflects the relative performance of NeRF-based neural rendering and Gaussian-based reconstruction methods. A clarification of the relationship between NeRF and Instant-NGP has been added to the revised manuscript (Line 536).

Line 104-106, The authors should more explain this sentence ‘’Compared with SfM-MVS, 3DGS also requires an initial sparse reconstruction via SfM; however, its key distinction lies in the sub- sequent dense representation stage’’.

RESPONSE: Thank you for your valuable suggestion. We have revised the manuscript to clarify the distinction between SfM–MVS and 3DGS. Specifically, both methods share a similar initial phase where Structure from Motion (SfM) provides camera poses and sparse 3D points. However, the subsequent stages are fundamentally different. MVS (Multi-View Stereo) generates a dense point cloud through pixel-level stereo correspondence, whereas 3D Gaussian Splatting (3DGS) constructs a continuous scene representation using Gaussian primitives that encode position, scale, color, and opacity parameters. These Gaussians are iteratively optimized via differentiable rendering to achieve photorealistic reconstructions. This difference allows 3DGS to produce smoother surfaces and higher rendering quality than traditional SfM–MVS. This revision is located at line 114 in the attached file.

Line 159-161: The conversion formula to HSV is not clear. Any references!

RESPONSE: Thank you for pointing this out. The RGB-to-HSV conversion in this study follows the standard formulation described in the reference “Enhancing Color Selection in HSV Color Space.” The citation has been added to the revised manuscript to clarify the source of the conversion method. line 191

4.In the data acquisition phase, how the authors extract randomly key frames from the video in the first step?

RESPONSE: Thank you for your valuable comment. We have clarified the frame extraction process in the revised manuscript. In this study, video frames were extracted at fixed temporal intervals using FFmpeg to obtain uniformly distributed viewpoints along the video sequence. Specifically, the output rate was set to fps = 2, which ensured even temporal coverage and minimized redundancy among consecutive frames, providing suitable multi-view inputs for 3D reconstruction. The corresponding description has been added to Line 182 in the revised manuscript.

5.What is the meaning of ‘’stochastic gradient descent’’?

RESPONSE: We thank the reviewer for pointing this out. Stochastic gradient descent (SGD) is an iterative optimization method widely used in machine learning. It updates the model parameters step by step to gradually minimize the difference between the rendered images and the real images. We have clarified this in the revised manuscript to improve readability for all readers. Line 282

6.Equation 8: authors should explain the meaning of ‘’T’’.

RESPONSE: We thank the reviewer for pointing this out. In Equation 8, The superscript T indicates the transpose operation, which converts a column vector into a row vector to enable valid matrix multiplication with the covariance components.. We have added this clarification in the revised manuscript to improve readability and ensure that the meaning of T is clear. Line 254

7.The authors should explain how they choose the calibration board?

RESPONSE: We thank the reviewer for the comment. We have clarified in the revised manuscript that a 10 cm × 10 cm calibration board with 6 mm × 6 mm black squares was used. This board provides sufficient reference points for accurate scale correction while being conveniently sized for seed imaging. Line 205

8.The authors should more explain how ‘’Alpha Shape’’ is analogous to the ‘’2D rolling ball method’’.

RESPONSE: We thank the reviewer for the comment. We have clarified the analogy between the Alpha Shape and the 2D rolling ball method in the revised manuscript. In 2D, a circle of radius α rolls along the point cloud, forming edges when it touches two points, whereas in 3D, a sphere of radius α forms triangles when it touches three points. This explanation has been added to improve clarity. Line 387

9.Extensive English language editing is needed.

RESPONSE: We appreciate the reviewer’s suggestion. The entire manuscript has undergone professional English editing through MDPI’s official Author Services to ensure grammatical accuracy, clarity, and consistency with the journal’s language standards.

10.Some abbreviations in the abstract such as SFM; 3DG should be defined.

11.Line 76: Reference ‘’Zermas et al.’’ should be numbered. The same point in line 95 (Mildenhall et al.

RESPONSE: Thank you for pointing this out. The in-text citations for Zermas et al. and Mildenhall et al. have been corrected, and their corresponding reference numbers have been properly inserted in the revised manuscript.

12.Species should be written in italic. Please check the whole manuscript.

RESPONSE: Thank you for your careful observation. All species names throughout the manuscript have been checked and formatted in italics according to the journal’s style requirements.

13.Please refer to the journal instructions when you written the different formula.

RESPONSE: Thank you for your reminder. All formulas have been carefully checked and reformatted according to the MDPI journal formatting guidelines, ensuring consistency in notation, numbering, and presentation throughout the manuscript.

14.The quality of figure 3 should be improved.

The legend of figures 4 and 5 should be more detailed.

Please check the legend of figure 6.

The figure 10 should be placed after the paragraph of line 484-488.

RESPONSE: Thank you for these valuable suggestions. Figure 3 has been remade to provide a clearer illustration of the 3DGS processing workflow.The legends of Figures 4–6 have been revised to provide more detailed and accurate descriptions. In addition, Figure 10 has been relocated to immediately follow the paragraph corresponding to lines 484–488 in the revised manuscript, as suggested.

15.Many punctuations and typos errors. Please check the whole manuscript.

References list should be checked and revised accordingly to the journal instructions (see references 28; 30; 32; 33).

RESPONSE: The reference list has been completely reformatted according to the official MDPI reference style (MDPI.ens) downloaded from https://www.mdpi.com/authors/references.