Machine Learning in the Design Decision-Making of Traditional Garden Space Renewal: A Case Study of the Classical Gardens of Jiangnan

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper focuses on the spatial renewal design decision support of Jiangnan classical gardens, applies the conditional generative adversarial network (CGAN) in artificial intelligence to the generation of traditional garden space layout, and attempts to solve the current situation of "heavy reliance on traditional experience and low design efficiency". The topic has practical urgency. In the renewal of urban historical blocks, the protection and reuse of traditional gardens are the core issues, responding to the dual direction of cultural heritage protection and technological empowerment. However, there are several problems:

 

(1) There are existing studies on the application of CGAN in urban green space generation, landscape layout, etc., and the article lacks a comparative study of existing similar methods (such as Diffusion models and Transformer models). CGAN technology itself is not innovative, but an application of existing technology in a specific field.

 

(2) It is recommended to increase the analysis of the similarities and differences between traditional designers' experience and model output, and strengthen the argument of "substitution" and "assistance". For example, add the subjective scoring of the generated scheme by professional garden designers.

 

(3) In terms of the rigor of the paper, the sample construction method has limitations: the original sample is only 142, and the 1,381 samples after expansion are still insufficient; a large number of samples are generated by rotation and mirroring, which poses the risk of "overfitting" and "pseudo-diversity".

 

(4) The evaluation of the model mainly relies on pixel-level image similarity, lacking structural or semantic accuracy verification.

 

(5) No open source code or reproduction instructions are provided, and reproducible code should be provided.

Author Response

Comments 1: There are existing studies on the application of CGAN in urban green space generation, landscape layout, etc., and the article lacks a comparative study of existing similar methods (such as Diffusion models and Transformer models). CGAN technology itself is not innovative, but an application of existing technology in a specific field.

Response 1:

We appreciate the reviewer’s insightful comment regarding the necessity of comparative analysis. While the CGAN is indeed an established technology, our work focuses on its domain-specific optimization for garden plane generation—a niche, yet critical subfield of landscape architecture where the CGAN’s advantages align uniquely with task demands.

Comparison of Generative Models and Applicability of CGAN:

Diffusion models generate high-quality samples through iterative denoising but incur high computational costs. Transformers excel at capturing long-range dependencies, yet their computational complexity escalates quadratically with data volume. CGAN, based on adversarial training, enables rapid generation (although it may face mode collapse issues).

For garden floor plan generation tasks, CGAN demonstrates distinct advantages:

• Its convolutional architecture efficiently captures local spatial features.
• Its generation speed meets real-time requirements for design refinement.
• Thus, the CGAN remains the optimal choice for balancing efficiency and output quality in current applications, being particularly effective for rapid iterative landscape design scenarios.

 

Comments 2: It is recommended to increase the analysis of the similarities and differences between traditional designers' experience and model output, and strengthen the argument of "substitution" and "assistance". For example, add the subjective scoring of the generated scheme by professional garden designers.

Response 2:

We sincerely appreciate your valuable suggestions. In response to your specific recommendation regarding incorporating professional designers' subjective evaluations, we have added a dedicated questionnaire assessment in the revised manuscript, collecting subjective ratings from professional landscape designers based on CGAN-generated solutions.

Methodology:

Four representative garden space typologies were selected, each providing one comparative set (Original solution vs. CGAN-generated solution), yielding eight total samples. Under double-blind testing, designers/decision-makers evaluated satisfaction across five dimensions:

1. Rationality of the layout;
2. Adaptability of the scale;
3. Richness of the content;
4. Aesthetic quality of the landscape;
5. Sense of landscape layering.

Key Findings:

• CGAN-generated solutions demonstrated significantly higher overall satisfaction than the baselines.
• Water-island Spaces and Pavilion Corridor Spaces achieved the highest satisfaction, with notable advantages in scale adaptability, content richness, and layering perception, highlighting the CGAN’s potential for spatial proportion optimization.
• Limitations:

• Rockery Sculpture Spaces:CGAN solutions showed slightly lower satisfaction than the original designs (particularly in layout rationality and aesthetics), indicating challenges in simulating complex natural forms.
• Botanical Spaces:Received the lowest evaluations, revealing limitations in botanical diversity and visual appeal.

These findings provide clear direction for future optimization: our research will prioritize refining the layout logic for rockery sculpture spaces and enhancing content richness in botanical spaces, while maintaining CGAN's technical strengths in water-island and pavilion corridor spaces to advance AI applications in landscape planning.

 

Comments 3: In terms of the rigor of the paper, the sample construction method has limitations: the original sample is only 142, and the 1,381 samples after expansion are still insufficient; a large number of samples are generated by rotation and mirroring, which poses the risk of "overfitting" and "pseudo-diversity".

Response 3:

We fully concur with your observation. We appreciate your concerns regarding the sample size. While the expansion from 142 original samples presents limitations, the scarcity of authentic classical garden cases necessitated geometric augmentation (rotation/mirroring) to enhance training stability. Critically, these transformations adhere to the principle of physical invariance in green space structures—preserving spatial topology and proportional relationships.

Specifically:

1. Spatial Feature Invariance: Rotational/mirror operations alter only orientation angles while maintaining fundamental spatial topology (e.g., layout logic of architecture-water-vegetation relationships), maintaining the validity of augmented samples for pattern learning.
2. Overfitting Control: Model convergence was ensured through real-time monitoring of training dynamics (balanced loss value fluctuations and output consistency across iterations).

 

Comments 4: The evaluation of the model mainly relies on pixel-level image similarity, lacking structural or semantic accuracy verification.

Response 4:

We sincerely appreciate the reviewer’s insightful suggestion regarding structural/semantic verification. In response, we have incorporated quantitative semantic accuracy analysis into the revised manuscript (Section 3.2), focusing on the fidelity of landscape element composition in CGAN-generated plans.

Key evidence from the analysis:

1. Element-Type Consistency:

The CGAN’s outputs preserved identical landscape element categories (e.g., water, vegetation, pavilions) compared to original designs across all four spatial typologies.

1. Minimal Proportional Deviation:

The pixel-area proportions of elements exhibited negligible differences between original and CGAN-generated plans:

• Water-Island Space: ≤0.08%deviation across 10 element types (e.g., Shrub: 3.99% vs. 3.95% [Δ=0.04%]; Water: Δ=0.02%);
• Rockery Sculpture Space: ≤0.25%deviation (4 elements);
• Pavilion Corridor Space: ≤1.44%deviation (9 elements);
• Botanical Spaces: ≤1.30%deviation (8 elements).

These results confirm that the CGAN achieves high semantic accuracy in replicating critical landscape components at the composition level, addressing the core concern of element-level authenticity.

 

Comments 5: No open source code or reproduction instructions are provided, and reproducible code should be provided.

Response 5:

The description of the open-source code is attached in the Appendix. The code related to the CGAN model has been uploaded to Supplementary File.

The full text has undergone English language editing by MDPI. The text has been checked for correct use of grammar and common technical terms, and edited to a level suitable for reporting research in a scholarly journal.

Reviewer 2 Report

Comments and Suggestions for Authors

Here are some suggested comments to improve the quality of the manuscript (buildings-3691979-peer-review-v1):

1. The transition from the research background to the problem statement could be smoother in the Introduction section. Explicitly link the challenges of traditional methods to the proposed CGAN solution earlier. Describe the accomplishments of previous studies, identify gaps, and outline your contributions accordingly.
2. Clearly explain the selection process for the 142 initial samples, including the criteria for representativeness and the exclusion of outliers.
3. More carefully proofread the entire manuscript to improve readability. Sections 4.1 and 4.2 present some repeated ideas; authors should avoid redundancy.
4. Explain the selection of rotation/mirroring over other augmentation techniques (such as scaling and noise addition) to enhance the manuscript's quality.
5. Add a table to summarize the key inputs and parameters used in the proposed Conditional Generative Adversarial Network (CGAN) model.
6. Provide more details on hyperparameters and the rationale for selecting 500 epochs.
7. Explain the methods used to calculate the 91.08% accuracy. In addition, including mathematical equations for the proposed model is essential for more reliability.
8. Examine potential biases in the test samples, such as their exclusion from training.
9. Acknowledge limitations in the bridge generation (Water-Island Space) and propose solutions (e.g., adding more bridge examples to training data).
10. Discuss why Pavilion Corridor Space showed slower convergence (e.g., complexity of architectural features).
11. Quantify time savings more vividly (e.g., "Reduced from 3 months to 10 seconds").
12. Enhance the quality of all figures by using high-resolution formats to ensure they are clearly visualized. Improve Figure 6 (iterative process) by adding labels to epochs (e.g., "Blurry contours at Epoch 10") for clarity. In Figure 7 (similarity comparison), include a scale bar or legend for pixel differences.
13. The article lacks a detailed discussion of the potential advantages, disadvantages, and limitations of the present methodologies used in the study, which could have provided a more comprehensive understanding of the research findings. The conclusion could benefit from a brief mention of potential future directions. Are there other modifications or additional features that could further improve the study's generalizability and applicability?

Author Response

Comments 1: The transition from the research background to the problem statement could be smoother in the Introduction section. Explicitly link the challenges of traditional methods to the proposed CGAN solution earlier. Describe the accomplishments of previous studies, identify gaps, and outline your contributions accordingly.

Response 1:

We sincerely appreciate the reviewer’s valuable guidance on enhancing the logical flow of the Introduction. We have fully addressed this concern through targeted revisions in Sections 1.2–1.3:

• Explicit machine learning linkage in Section 1.2

At the beginning of 1.2, the current application status of machine learning in the research of traditional garden landscapes is proposed. Subsequently, the corresponding research results were specifically detailed.

“In recent years, researchers have begun to adopt machine learning methods to quantitatively analyze the characteristics of garden spaces. The rapid development of machine learning technology has provided new solutions for this field.”

• Clear gap identification in Section 1.2

At the end of 1.2, the deficiencies of current machine learning approaches for the generation of garden plans have been added. This leads to the corresponding contribution of this article in 1.3.

“However, existing CGAN applications primarily focus on generic urban green spaces, with two domain-specific gaps persisting in machine learning implementations: (1) Specialized frameworks tailored to Jiangnan garden design principles; (2) Integration of fine-grained historical spatial metrics (element typologies, combinatorial patterns); (3) Imbalanced distribution of spatial element categories. These identified limitations constitute the primary directions for future research.”

• Restructured Section 1.3 content:

Aiming at problems including traditional garden design’s high dependency on experience, prolonged scheme generation cycles, and existing deficiencies in machine learning applications within this domain, this study proposes a CGAN-based spatial analysis and generation framework for Jiangnan gardens. The characteristics of this framework, primary research questions, and contributions are presented.

The full text has undergone English language editing by MDPI. The text has been checked for correct use of grammar and common technical terms, and edited to a level suitable for reporting research in a scholarly journal.

 

Comments 2: Clearly explain the selection process for the 142 initial samples, including the criteria for representativeness and the exclusion of outliers.

Response 2:

The criteria for selecting initial samples have been supplemented in the manuscript:

The screening criteria for initial samples are:

(1) Area selection: Between 200 and 1000 m², to control the scale rationality of spatial generation. Excluded are instances like the 8 m² plant courtyard in the Lingering Garden.

(2) Scope definition: Possess relatively clear spatial boundaries. For example, the scope of architectural spaces is defined by building boundaries, and plant spaces are defined by plant courtyards.

(3) Distinct element characteristics: Each spatial type must contain clearly identifiable landscape elements. For instance, a water-islet space must include water bodies, and a rockery space must contain rockery. Excluded are spaces dominated by a single element, such as a 200 m² plant space filled solely with bamboo, or an area consisting entirely of water bodies without other landscape elements.

 

Comments 3: More carefully proofread the entire manuscript to improve readability. Sections 4.1 and 4.2 present some repeated ideas; authors should avoid redundancy.

Response 3:

We thank the reviewer for their careful reading and valuable suggestions regarding the manuscript’s readability and redundancy. We have carefully implemented the following revisions throughout the manuscript:

Comprehensive Proofreading: The entire manuscript has been meticulously proofread to enhance clarity, flow, and overall readability.

Addressing Redundancy in Sections 4.1 & 4.2: We have specifically revised Sections 4.1 and 4.2 to eliminate repeated ideas and ensure conciseness.

Section 4.1: The discussion concerning the requirement for deep design expertise and the lengthy nature of the traditional design cycle has been restructured to remove redundant descriptions and improve focus.

Section 4.2: Redundant text identified in this section has been removed to avoid repetition and strengthen the presentation of distinct points.

 

Comments 4: Explain the selection of rotation/mirroring over other augmentation techniques (such as scaling and noise addition) to enhance the manuscript's quality.

Response 4:

We thank the reviewer for their query on our augmentation choices. Rotation and mirroring were specifically selected over alternatives (such as scaling, noise addition) to preserve core physical invariants in classical garden design:

1. Preserves Critical Invariants: These operations maintain absolute scale, proportional ratios, and spatial topology (e.g., element adjacency, sequence). Scaling distorts proportions; noise addition creates implausible layouts, violating design principles.
2. Ensures Valid Augmentation: By upholding these invariants, augmented samples remain authentic representations of the spatial patterns for model learning.

Therefore, rotation/mirroring uniquely balanced dataset growth with strict adherence to the spatial truths of classical gardens.

 

Comments 5: Add a table to summarize the key inputs and parameters used in the proposed Conditional Generative Adversarial Network (CGAN) model.

Response 5:

We thank the reviewer for this constructive suggestion. We have supplemented the key parameter tables and explanations used in the training of the CGAN model in 3.1. The inclusion of this table significantly enhances methodological transparency and ensures the reproducibility of our experiments.

 

Comments 6: Provide more details on hyperparameters and the rationale for selecting 500 epochs.

Response 6:

Thank you for your valuable suggestions. The rationale for setting 500 training epochs is twofold:

 (1) Past experience:

Based on previous training experience, it is generally possible to complete the training process by setting 300 to 500 generations of training. Due to the complexity of the garden landscape space, the researchers chose 500 generations of training to ensure the effectiveness of machine learning.

(2) Loss Convergence Evidence:

As demonstrated in Figure 4 (Section 3.1.1), the loss curves of all four model variants exhibited a sustained decline throughout training. Critically, after ~480 epochs, the losses stabilized within a low-value regime (<1.0) with minimal fluctuations (±0.05), indicating convergence of both discriminative and generative components. This empirically confirms that 500 epochs provide:

Model Stability: Output consistency across final training iterations (Epochs 480–500),

Performance Saturation: No further accuracy gains observed beyond 490 epochs.

The epoch count was intentionally set slightly above the observed convergence point (480 epochs) to ensure robust optimization margins.

 

Comments 7: Explain the methods used to calculate the 91.08% accuracy. In addition, including mathematical equations for the proposed model is essential for more reliability.

Response 7:

We wrote a Python code to calculate the pixel similarity of images. This piece of code mainly uses the basic operations in image processing and does not involve complex mathematical formulas or algorithms. The core mathematical operations are simple matrix/pixel-level operations (subtraction, absolute value, threshold comparison) and percentage calculations.

We conducted similarity calculations on the images generated by the four groups of models after learning 480 generations (model convergence), and obtained an average similarity of 91.08%.

The above content has been supplemented and explained in 3.2.

 

Comments 8: Examine potential biases in the test samples, such as their exclusion from training.

Response 8:

To test the generation ability of the model, the test samples chose spatial contours that were completely different from the original samples, such as randomly sketched irregular figures, regular combinations of rectangles, circular spaces, etc. These graphics did not appear in the corresponding training process.

 

Comments 9: Acknowledge limitations in the bridge generation (Water-Island Space) and propose solutions (e.g., adding more bridge examples to training data).

Response 9:

Bridge elements are not universally present in every classical garden. The relatively few samples of bridge-containing spaces in existing traditional gardens have resulted in less mature outcomes for CGAN-generated Water-island Spaces. In future research, we will collect more water-island spatial samples with bridges to enhance CGAN training.

 

Comments 10: Discuss why Pavilion Corridor Space showed slower convergence (e.g., complexity of architectural features).

Response 10:

Thank you for highlighting this issue.

The Pavilion-Corridor Spaces exhibited slower convergence due to their heightened architectural–vegetal complexity. The model encountered optimization challenges when processing:

Intricate architectural contours (curved eaves, interconnected pavilion-corridor rooflines)

Multi-scale spatial dependencies (vegetation penetration through corridor frameworks)

These complexities induced gradient instability during training, manifesting as significant loss fluctuations and protracted convergence.

We have discussed these issues as supplementary text in Section 3.1.1.

 

Comments 11: Quantify time savings more vividly (e.g., "Reduced from 3 months to 10 seconds").

Response 11:

We thank the reviewer for this valuable suggestion.

In Section 4.1, we have highlighted the quantitative description of time savings and adopted the vivid contrast format suggested by the reviewer.

 

Comments 12: Enhance the quality of all figures by using high-resolution formats to ensure they are clearly visualized. Improve Figure 6 (iterative process) by adding labels to epochs (e.g., "Blurry contours at Epoch 10") for clarity. In Figure 7 (similarity comparison), include a scale bar or legend for pixel differences.

Response 12:

We appreciate your suggestion.

Regarding the blurriness observed in Generation 10 (Figure 6):

Image resolution during training correlates directly with the model's learning progress.

Early iterations typically yield lower-resolution outputs (appearing blurry), while increasing epochs gradually enhances clarity as feature extraction refines.

Thus, the Generation 10 output represents an authentic early-stage result, not an artifact of low source resolution.

Additionally, Figure 7 now includes a corresponding pixel-difference legend to provide clearer visualization of matching/non-matching regions.

 

Comments 13: The article lacks a detailed discussion of the potential advantages, disadvantages, and limitations of the present methodologies used in the study, which could have provided a more comprehensive understanding of the research findings. The conclusion could benefit from a brief mention of potential future directions. Are there other modifications or additional features that could further improve the study's generalizability and applicability?

Response 13:

The selection of Model 2.2.1 has been supplementarily discussed in the text, and we have specifically explained the advantages of choosing the CGAN model as the generation tool in this paper.

The future scalable research directions have been supplemented in the Conclusion. For example, testing the model performance against a high-density urban background to make the model applicable to more complex urban heritage spaces. In the future, functional expansions from planar generation to 3D model generation and other contents are planned.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The author answered the questions very well and I have no further questions.

Reviewer 2 Report

Comments and Suggestions for Authors

Accept in present form.