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

Abstract

This research takes the Suzhou Gardens, a World Cultural Heritage Site, as the object of study and investigates a rapid scheme generation approach for garden restoration and expansion projects, assisting designers in making scientific decisions. Considering the limitations of current garden design, which is inefficient and relies on human experience, this study proposes an intelligent generation framework based on a conditional generative adversarial network (CGAN). In constructing the CGAN model, we determine the spatial characteristics of the Suzhou Gardens and, combined with historical floor plan data, train the network. We then design an optimization strategy for the model training process and finally test and verify the generative space scheme. The research results indicate the following: (1) The CGAN model can effectively capture the key elements of the garden space and generate a planar scheme that conforms to the traditional space with an accuracy rate reaching 91.08%. (2) This model can be applied to projects ranging from 200 to 1000 square meters. The generated results can provide multiple scheme comparisons for update decisions, helping managers to efficiently select the optimal solution. (3) Decision-makers can conduct space utilization analyses and evaluations based on the generated results. In conclusion, this study can help decision-makers to efficiently generate and evaluate the feasibility of different design schemes, providing intelligent support for decision-making in urban renewal plans.

1. Introduction

1.1. Research Background

Urban renewal is crucial for revitalizing historic cities, and effective site planning is key to ensuring successful redevelopment [1]. In recent years, local governments have paid increasing attention to the renewal of landscape spaces within historic conservation areas. Spatial design is crucial for the spatial renewal of these historically protected areas. The traditional gardens of Jiangnan are outstanding representatives of Chinese classical garden art and an important part of the world’s cultural heritage [2].

However, the acceleration of urbanization and changing social demands are challenging the protection and renewal of Jiangnan gardens. Within Suzhou’s historic and cultural city conservation area, conducting spatial renewal while preserving its historical features has become a critical issue for decision-makers (Figure 1).

At present, decisions regarding the renewal design of Jiangnan gardens still rely on the experience and judgment of designers, lacking systematic and scientific methodological support. Most urban renewal decision-making methods often result in homogeneous spaces that lack local characteristics [3]. The traditional garden design process is often time-consuming and fails to fully consider the complex characteristics of the space. The challenge for designers lies in effectively preserving characteristics such as spatial scale, types of landscape elements, rules of landscape combination, and the layout of landscape space. Furthermore, due to the lack of quantitative analysis tools, many renewal plans fail to fully conserve the spatial essence of Jiangnan gardens, resulting in unreasonable design decisions and possibly even destruction of the original cultural values and landscape continuity.

Therefore, efficiently and accurately extracting and applying the characteristics of garden spaces in the renewal of historic protected areas has become a key issue in improving the efficiency and quality of design decisions.

1.2. Literature Review

Researchers have recently 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 in this field. Some scholars have applied computer vision technology to the recognition of spatial types; for example, the intelligent recognition and classification of the types of images of architectural ruins in Jiangnan were achieved using machine learning [4,5]. Intelligent recognition of surface damage in Jiangnan garden architecture was performed based on the computer vision technology of machine learning [6]. An analysis framework for spatial indicators in Chinese classical gardens has been constructed and studied based on spatial syntax and machine learning [7]. The modular information fusion model of garden landscapes was studied using machine learning algorithms [8]. Research on the characteristics of traditional garden roads and the intelligent generation of space, etc., has also been conducted [9]. In terms of spatial layout, some scholars have studied the layout characteristics of the geometric space of Jiangnan gardens using machine learning [10,11]. In particular, the application of GAN technology in landscape space has gradually matured. For example, the conditional generative adversarial network (CGAN) was applied to the generation of landscape plan layouts [12,13], and GAN technology has been used for the determination of spatial area and prediction of garden landscapes, etc. [14].

In terms of design decision-making, integrating machine learning technology into the garden decision-making process is shaping the current research frontier. In recent years, scholars have focused on the application of artificial intelligence in landscape space design and management, such as garden planning systems, site selection, and spatial quality assessment [15]. For example, intelligent planning and design were achieved through the integration of multi-source data [16,17]. The site selection of landscape space was accomplished using machine learning [18]. Research has been conducted on spatial renewal, transformation design, quality assessment based on machine learning, etc. [19,20,21].

In conclusion, the application of machine learning technology in the landscape renewal design process has brought about major breakthroughs in this field. By relying on the power of automated data processing methods, the efficiency of design analysis can be significantly improved. In particular, the conditional generative adversarial network (CGAN) can automatically generate garden layouts. However, existing CGAN applications primarily focus on generic urban green spaces, with several domain-specific gaps persisting in machine learning implementations: (1) there is no specialized framework for the design principles of Jiangnan gardens; (2) they lack the integration of small-scale historical spatial scales (element types, combinations); (3) and the distribution of spatial element categories is imbalanced. These limitations constitute the primary directions for future research.

1.3. Problem Statement and Objectives

This study considers the renewal of historically protected areas in Jiangnan gardens as a case study. Addressing the limitations of traditional garden design—including its heavy reliance on empirical expertise, prolonged scheme generation cycles, and existing gaps in machine learning applications for this domain—we pioneer a CGAN-based framework for the spatial analysis and generation of Jiangnan gardens. Leveraging CGAN technology to decode critical spatial patterns, our approach enables rapid scheme generation for garden restoration/expansion, effectively streamlining design decision-making timelines. This study aims to explore the application of machine learning technology in the decision-making process regarding the spatial renewal design of traditional gardens in the Jiangnan region. We construct a data-driven auxiliary design framework, with the expectation of providing decision support for the protection and innovation of world cultural heritage.

Our objectives were as follows:

(1)

Rapid generation of typologically diverse schemes adhering to spatial characteristics;

(2)

Preservation of landscape features at fine-grained heritage spatial scales;

(3)

Quantitative evaluation of spatial element compositional ratios.

By harnessing the ability of CGANs to decode complex spatial relationships, our approach uncovers latent patterns beyond the reach of traditional methods. Furthermore, at the design decision-making level, it facilitates a paradigm shift from empirical approaches to data-driven science in landscape research, delivering robust technical support for the contemporary reinterpretation and design decisions of historic spaces.

Through our planning practice and research, the following six issues are mainly discussed:

(1)

What types of landscape spaces exist in traditional Jiangnan gardens?

(2)

What are the specific methods and processes of CGAN technology in the generation of traditional landscape spaces?

(3)

How accurate are the data training process and test results?

(4)

Compared with the traditional design process, what advantages does the process of generating floor plans with CGAN have in design decision-making?

(5)

Can multiple schemes be output to meet the decision-makers’ needs for scheme comparison and selection?

(6)

Does the utilization rate of the generated planar space meet the normative requirements of the design decision?

2. Materials and Methods

2.1. The Characteristics of Garden Space

Traditional Jiangnan gardens, renowned for their poetic integration of architecture, nature, and philosophy, employ meticulously designed landscape spaces to create harmony and layered visual experiences [22,23]. Generally, based on the characteristics of landscape space elements, the structure of their main spatial types can be classified into the following types [24,25,26].

(1)

Water-Island Space

This is a space with water as its scenery; for example, central ponds, rivers, and streams. The curved waterways guide the movement, with zigzag bridges amplifying spatial depth.

(2)

Rockery Sculpture Space

The garden mainly features artificial mountains, using porous Taihu Lake rocks to simulate peaks or mountain ranges; for example, Lion Grove Garden’s labyrinthine rock formations.

(3)

Pavilion Corridor Space

The space is composed of pavilions, terraces, and towers in the garden; for example, open-sided structures for contemplation. The corridor space that winds through the garden serves as a walkway and spatial divider.

(4)

Botanical Spaces

These usually include floral focus areas, such as lotus ponds in summer, osmanthus groves in autumn, and prunus mume (plum) clusters in winter, etc. Also present are bamboo groves planted near walls to cast moving shadows, and tall trees (such as ginkgo) providing shade, mid-layer shrubs (camellias), and ground covers like moss or ferns (Figure 2).

Traditional Jiangnan gardens artfully integrate aquatic, rockery, architectural, and botanical spaces to create harmonious, layered landscapes that blend nature with poetic philosophy [27]. This study respects the traditional classification of spatial types and employs these four spatial types: rockery sculpture space, water-island space, pavilion gallery space, and botanical spaces.

2.2. CGAN Method

2.2.1. Framework and Principles

This study adopts a CGAN (conditional generative adversarial network) as the core framework for garden floor plan generation. CGAN is an extension of GAN (generative adversarial network), and it is mainly applied in image generation tasks [28].

Compared to CGAN, diffusion models and Transformer models exhibit distinct advantages and limitations in generative tasks, primarily differing in generative mechanisms, training stability, and application scenarios. Diffusion models produce high-quality samples through stepwise denoising, excelling in output diversity but suffering from high computational costs and slow generation speeds [29]. Transformer models rely on self-attention mechanisms to achieve the global modeling of structured data, yet their computational complexity increases quadratically with input size, and they are prone to overfitting on small datasets [30]. CGANs enable rapid generation via adversarial training, but face issues such as mode collapse and training instability in complex scenarios [31] (

Table 1. Key technical comparisons.

Model Type	Strengths	Limitations	Suitability for Garden Floor Plans
Diffusion Models	High-quality details	High computational cost; slow generation	Limited (real-time iteration required)
Transformers	Global context modeling	Quadratic complexity; data-hungry	Overkill (local patterns > global links)
CGAN	Fast generation; local feature capture	Mode collapse risks	Optimal: Balances speed and spatial granularity

).

CGAN remains the optimal methodological framework due to three practical advantages: (1) single-pass inference enables the real-time generation essential for design iteration, outperforming diffusion models in speed; (2) convolutional inductive bias efficiently captures the local spatial patterns critical to garden layouts, avoiding the computational overhead of Transformer models; (3) architectural priors ensure robustness with limited data and reduce overfitting risks, which balances speed, feature localization, and adaptability in resource-constrained scenarios. Therefore, the CGAN was selected as the methodological framework for this study, offering the optimal balance of speed, localized feature capture, and practical adaptability within resource-constrained scenarios.

The CGAN framework consists of a generator and a discriminator (Figure 3). Among them, the function of the generator is to generate images, while that of the discriminator is to determine whether the generated images are real images. Unlike ordinary GANs, CGANs require that additional conditions be provided, and these conditions can be images of any form [32]. By introducing conditions (TrainA), CGAN can generate more targeted and personalized images.

During the CGAN training process, the generator and discriminator are trained through adversarial learning. The generator deceives the discriminator (TrainB) by learning how to generate images that match the condition (TrainA), while the discriminator prevents deception by learning to distinguish between the real images and generated images [33]. During this cyclic training process, the generator and the discriminator keep iterating until the generator can generate sufficiently realistic images, such that the discriminator is unable to determine which images are real. After the training is completed, by adjusting the input conditions, the attributes of the generated images can be controlled, thereby obtaining images with specific features. This study, based on the principle of the CGAN generative adversarial network model and taking the spatial characteristics of different garden artistic conceptions as conditions, generated garden floor plans with different landscape spatial characteristics.

The goal of this study is to enable the machine to automatically generate floor plans with the spatial characteristics of garden landscapes after learning a large amount of sample data through CGAN. This not only retains the planar characteristics of the traditional garden landscape space but also serves as an intelligent tool for generating garden schemes.

2.2.2. Research Methods and Procedures

The method of generating the spatial floor plan of garden landscapes using the CGAN comprises a new attempt to combine the protection of garden cultural heritage with AI. This method mainly consists of five process steps, namely, data collection, data processing, model training, model evaluation, and model application (Figure 4). The specific method contents are as follows.

(1)

Data classification: Spatial type division is carried out in accordance with traditional landscape features. According to the characteristics of landscape space elements in traditional Jiangnan classical gardens, their spatial types are classified as: rockery sculpture space, water-island space, pavilion corridor space, and botanical spaces.

(2)

Data collection: At the same scale, spatial floor plans with corresponding features are extracted from the original floor plans of classical gardens according to different types of landscapes as the basic samples for training.

(3)

Data processing: Targeted processing is conducted based on the characteristics of the samples. Firstly, the sample size is increased. The sample size of the landscape space is limited, and it is necessary to increase the training sample size through technical means. Secondly, the contour map is drawn, which will serve as an input for the generation condition in the CGAN learning process. Secondly, the landscape space elements are classified and colored to enable the machine to effectively recognize the content and features of the drawing during the learning process. Finally, the size of the training images is unified to improve the efficiency of training.

(4)

Model training: Paired training samples are placed in the CGAN training framework for machine learning. The CGAN model framework requires paired samples for training. The “contour atlas” and the corresponding “spatial element atlas” obtained above are paired sample sets. Each of the contour maps corresponds to a spatial element map for paired training. The purpose is to obtain the spatial meta-layout primitives under contour control; that is, to input one contour map and derive a corresponding spatial element layout map scheme.

(5)

Model Evaluation: To verify the training effect of the model, this study conducted evaluations in three aspects. First, the loss value of CGAN was observed in each iteration, and test images were generated during the iteration process. Secondly, the generated test samples were compared with the original samples to evaluate whether there was any deviation in the effect of the generated samples. Finally, a professional satisfaction survey was conducted between the samples generated by the test and the original samples to obtain a professional assessment of the spatial effect. After the above three rounds of evaluation, if the effect was not satisfactory, it was necessary to adopt cyclic optimization training, repeatedly adjusting parameters, increasing the number of iterations, and employing other methods to improve the effect of model training. If the training effect was satisfactory, the next application stage could be entered.

(6)

Model Application: The results generated through the training of this process method are applied in actual design as an auxiliary tool for the generation of garden landscape space.

2.3. Data Collection and Processing

The collection of floor plans: Based on the four types of spaces (rockery sculpture space, water-island space, pavilion corridor space, botanical spaces), a total of 142 floor plans of the corresponding landscape space types were selected from the classical garden floor plans. The screening criteria for the 142 initial samples were:

(1)

Area selection: Between 200 and 1000 m², to control the scale rationality of spatial generation. Excluded were instances like the 8 m² floor plan courtyard in the Lingering Garden.

(2)

Scope definition: It must 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 identifiable landscape elements. For instance, a water-island space must include water bodies, and a rockery space must contain rocks. Excluded were 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.

To ensure the proportion and scale of the original floor plan, the spatial range was cut at the same scale of 1:100. Among the characteristics were 27 mountain and rock types, 38 water feature types, 35 architectural types, and 42 flower and tree types. The size of the space obtained after interception ranged from 200 to 1000 m². The space obtained after training could be applied in the renewal of micro and small spaces in urban historic protection blocks.

Increase the sample size: Obviously, the 142 images were not sufficient for model training. Therefore, we expanded the sample size. By rotating the original samples by 90°, 45°, or mirroring them, the number of samples could be increased [34]. After each sample was rotated and mirrored in all directions, it could be extended to 6 to 10 samples. The garden floor plans after rotation and mirroring did not undergo any changes such as scaling, distortion, or deletion, nor was the original layout content of the floor plan altered. The final total sample size reached 1381, which met the training requirements of the samples. Among them, there were 235 mountain and rock types, 380 water feature types, 346 architectural types, and 420 flower and tree types.

Draw the spatial contour: After the training quantity was met, it was necessary to draw the contour of the sample. The specific operation filled the garden floor plan obtained by the above organization with black according to the outline to obtain the “outline atlas.” The contour map was used as the input condition of the CGAN.

Classification and coloring of spatial elements: To enable the machine to effectively read the complex information on the drawing surface, color separation processing was performed on the different landscape elements in the floor plan (one color represented one landscape element). According to the spatial elements of the landscape, a total of 12 color types were classified. The buildings are red (R255, G0, B0), the pavilions and structures are rose red (R255, G0, B255), the corridors are yellow (R255, G255, B0), the bridges are dark blue (R0, G0, B255), the water is light blue (R0, G255, B255), and the trees and shrubs are green (R0, G255, B0). The ground cover plants are dark green (R9, G124, B37), the garden paths are orange (R255, G150, B0), the rockery stones are brown (R106, G57, B0), the platforms are cyan (R0, G117, B169), and the open spaces in the square are pinkish green (R172, G213, B152). The revetments are light yellow (R245, G245, B156). The classification and annotation of colors not only help machines identify the differences in image content, but also improve the feature learning of machines for different types of landscape spaces [35]. After color-marking each floor plan, the “spatial element layout atlas” corresponding to the abovementioned “outline atlas” was obtained (Figure 5).

Unify image size: To avoid generation errors caused by differences in sample image sizes, the above two image sets were uniformly processed into 512 × 512 pixels, 96 dpi resolution, and 24-bit depth images.

3. Model Training and Testing

3.1. Model Training Process

The sorted paired samples were input into the CGAN for training (Appendix A). The “contour atlas” was the conditional sample, and the “spatial element layout atlas” was the generation sample to train the layout generation of spatial elements under the control of the contour conditions. To ensure the effectiveness of the training process, it was necessary to constantly observe the loss logs during the training process and the test images in the iterations.

Based on prior training experience, 300–500 epochs are typically sufficient to complete the training process [36]. Given the complexity of landscape garden spaces, 500 epochs were selected to ensure effective machine learning. To preserve the generic features learned during training (edges, textures) while significantly reducing computational load and memory consumption, freezing was implemented at epoch 150. Additionally, model snapshots were saved every 50 epochs to facilitate progress monitoring (Appendix B). This configuration ensured training validity, retained critical learned features, and enabled efficient process inspection (

Table 2. Key parameters of model training.

Parameter Type	Input
Picture size	512
Training Epochs (niter)	500
Freezing Epochs (niter decay)	150
Checkpoint Interval (save epoch freq)	50

).

3.1.1. Loss Value Log Data

During the model training process, loss logs were generated simultaneously, and the loss values in them reflected the training quality of the generative adversarial network [37]. A loss value with a continuous downward trend indicated that the samples generated by the model during the learning process were constantly approaching the real samples. A loss value no longer decreasing or fluctuating indicated that the learning of the model reached a bottleneck and an over-fitting situation occurred [38].

In this study, the corresponding loss value line graphs were obtained by statistically analyzing the loss values of each iteration of the generator and discriminator during the model training process of the four landscape types (Figure 4). As can be seen from Figure 5, the loss curves of all four model variants exhibited a sustained decline throughout training. Critically, after ~480 epochs, losses stabilized within a low-value regime (<1.0) with minimal fluctuations (±0.05), indicating convergence of both discriminative and generative components. The trained models were deemed effective.

Among them, the models of the water-island space (a) and the rockery sculpture space (b) had a relatively fast convergence speed, and the downward trend of the loss value curve was obvious. The reason is that the spatial form characteristics of mountains and water are relatively obvious, and there are significant differences among the various elements, so the efficiency of model learning is relatively high. Although the loss value curves of the pavilion corridor space (c) and the botanical spaces (d) also declined significantly, the convergence process of the models fluctuated greatly. Due to the complexity of the buildings and plant spaces, difficulties emerged in the learning process of the model, especially in the architectural space, the complex architectural contours (curved eaves, interconnected pavilions, corridors and roof lines), and the multi-scale spatial dependence (the infiltration of vegetation through the corridor frame). These complexities led to unstable gradients during the training process, causing significant fluctuations in the learning process and a slower convergence speed. Nevertheless, all four groups of models gradually converged after nearly 500 epochs and achieved good learning results in terms of data (Figure 6).

3.1.2. Accuracy of the Iterative Process Model

In addition to observing the model’s learning status from the loss value data, it was also necessary to constantly observe whether the test images during the iteration process gradually became clear and realistic, as well as whether there were any conditions such as blurriness or distortion. This allowed us to determine whether the model converged or whether the machine learning parameters needed to be readjusted.

During the model training process, one check image was set to be output every 10 epochs to determine changes in the model’s learning accuracy. As shown in Figure 7, the effect generated by the model in the 10th epoch is very blurry. By the 100th epoch, there is already a relatively clear spatial outline. Starting from the 200th epoch, the effect begins to approach the real picture. Until the 500th epoch, the effect is almost consistent with the real picture. This indicates that the model has basically achieved the training objective after 500 epochs of training. The results produced by the generator have learned the characteristics of these types of garden spaces (Figure 7).

In terms of image similarity, the training effects of the water-island space, the rockery sculpture space, and the botanical spaces were relatively good. In the 100th epoch, the spatial forms are quite different from the real pictures, and there were many errors starting from the 200th epoch; however, the results from the generator were very close to the real pictures. The morphological elements of the pavilion corridor space were rather complex, and the efficiency of the model in learning the spatial morphological rules was relatively slow. The results at 200 epochs still exhibited many errors when compared with the real pictures. However, with a continuous increase in the number of training sessions, the spatial generation effect of the architectural type was almost the same after 500 epochs of training as that of the real picture.

In summary, after 500 epochs of training by CGAN, the four sets of loss value log data for the spatial floor plan samples of the four landscape types finally began to converge at the 480th epoch. The results generated by its generator were almost consistent with the real samples. This indicated that training the model to generate the “layout diagram of landscape space elements” through “outline” was effective. The model completed the feature learning of the landscape space and had a good effect. The model training was completed.

3.2. Model Accuracy Test and Comparison

After the model was trained, it was further tested using new samples to evaluate its performance. After the new untrained samples were contoured, they were placed as input samples into the trained CGAN program. The generated results were compared with the original real samples to test the similarity and accuracy of the generated results. The testing comprised two components: shape contour similarity (assessed via image pixel matching) and spatial element semantic accuracy (measured by compositional deviation of element types).

As shown in Figure 8, after four groups of new landscape samples were placed in the trained CGAN model, the output results were found to be extremely similar to those of the real samples. For shape contour similarity, a Python 3.8.3 script was developed to calculate image pixel matching. By calculating the pixel differences between the two images, the specific locations and similarity percentages of the morphological differences between the two were obtained more intuitively.

Among the spaces, the similarity of the water-island space was the highest, reaching 99.31%. The morphological differences occurred in the contour lines, and the differences were extremely small. The pavilion corridor space and botanical spaces showed similarities of 95.33% and 95.23%, respectively. The differences in pavilion corridor space were located at the positions of water areas and some roads, while the morphological differences in botanical spaces mainly lay in the plants and roads at the left and right ends. The overall morphological similarity was very high. In contrast, the morphological differences in the mountain and rock scenes were relatively greater, but the similarity still reached 94.07%. The main difference lay in the treatment of the contours of the mountain and rocks (Figure 8). Based on the training logs, similarity analysis was conducted on the images generated by the four model groups at epoch 480 (model convergence), yielding an overall average similarity of 91.08%.

The accuracy of the landscape semantics of the floor plan generated by CGAN and the original floor plan was determined by calculating the color proportions of landscape element types. As shown in Figure 9, among the four groups of typical plans, the types and quantities of elements generated by CGAN were consistent with the original floor plan types.

Among the 10 spatial types in the first group of water-island spaces, the difference between the elements generated by CGAN and the original image ranged from 0.00% to 0.08%. Among them, the shrub element generated by CGAN was 3.99%, which was 0.04% higher than the 3.95% of the original image; the tree was 0.03% different, the artificial rockery was 0.07% different; and the water area was 0.02% different. Garden road, square, and open space differed by 0.08%, while revetment, pavilions and structures, and platform differed by only 0.01%. Similarly, in the comparison of the four types of elements in the second group, rockery sculpture space, the difference in the proportion of elements was 0.00% to 0.25%. The comparison differences of the nine element types in the third group, pavilion corridor space, ranged from 0.00% to 1.44%. The comparison differences among the eight element types in the fourth group of botanical spaces ranged from 0.00% to 1.30% (Figure 9).

It can be seen from this that the proportion of planar element types generated by CGAN was extremely different from the original image. This reflects the relatively high accuracy of the CGAN generation floor plan in terms of element semantics.

3.3. Model Robustness Test

To verify whether the model can generate multiple schemes, conducting a robustness test of the CGAN model is an important step [39]. From the test model samples, one representative floor plan from each of the four groups of landscape space types was selected as the test object. When changing the noise input of the model, the four groups of landscape space models generated a variety of different morphological schemes.

As can be seen from Figure 10, the spatial elements generated by each group of models exhibited differences while conforming to the theme. Among them, the pavilion corridor space changed significantly. The reason is that there are many types of spatial elements around the building, especially the integration of water elements, which enrich this entire space. Among the different schemes for generating water features, the main characteristics of the water surface were retained, and the positions of the pavilions and garden paths were also changed accordingly. However, there were slight flaws in the generation of bridges in Effects 3 and 4 of the water-island space generation, indicating that the learning of bridges in water-island space required improvement. The test results of the water-island space showed a relatively stable state, and the generated planar forms were quite similar. The results of multiple schemes for botanical spaces varied significantly. Pavilions were generated in Effects 1 and 2, while no pavilions were generated in Effects 3 and 4 (Figure 10). These locally occurring problems can be solved by increasing the sample size or the number of training sessions.

In summary, the robustness test results of the four groups of CGAN models were relatively ideal. The models generated multiple schemes that conformed to the topic from the same contour norm space. These generated schemes demonstrated rich creativity in various aspects, such as the proportion, position, form, and outline of spatial elements. This model can generate diverse design schemes or inspirations with the spatial characteristics of Jiangnan gardens for architects in different spatial scenarios, thus creating more flexible and personalized garden space forms.

Meanwhile, the author also tested the model’s generation of random contour samples. The randomly drawn planar contours were used as input conditions for the four groups of trained models to observe the generated effects. The results of the random generation experiment are shown in Figure 11. Under the randomly drawn outline, the model generated the element layout content corresponding to the topic.

In terms of water-island space generation, the model demonstrated stable performance within the contour range. It was capable of generating different water body forms that conformed to the size of the space. In terms of the generation of rockery sculpture space and botanical spaces, the model combined flowers and trees with mountain stones to generate various forms of mountain stone and flower and tree combination spaces. It automatically adjusted the proportion of the rockery and plants according to the rock, flower, or tree theme. In terms of pavilion corridor space generation, the generation effect of element shapes was relatively poor. However, the layout framework between the building and its various elements could still be observed from it. It formed an architectural space with a pavilion as the core, surrounded by trees or embraced by corridors. The problem of random generation results of architectural models can be solved by generating more schemes and selecting a more reasonable option (Figure 11).

Overall, after the robustness test of the model and the random contour sample experiment, we determined that the generated floor plans conformed to the basic effect of creating a garden landscape space. In the future, the clarity of architectural forms must be further optimized. These test results prove the reliability and practical value of the model. It is capable of effectively generating layout schemes that meet the morphological characteristics of various landscape spaces.

4. Discussion: Application in Design Decision-Making

4.1. Simplify the Decision-Making and Design Process of the Scheme

The well-trained CGAN has shown significant application potential in the field of landscape floor planning and design. With the help of the trained CGAN model, it can assist designers or decision-makers in quickly generating planar schemes that conform to the local garden spatial structure, reducing the scheme design time.

The design and renovation of the Suzhou Gardens Conservation Area, a World Cultural Heritage site, demand that designers possess extensive professional knowledge and design experience. Designers or decision-makers require time to evaluate the characteristics of traditional garden spaces and must transform this knowledge into practical design experience. During the design process, preliminary preparations must be integrated with the extraction of traditional spatial features, and the scheme undergoes multiple revisions before finalization. Extracting these traditional spatial features necessitates a deep understanding of and rich experience with spatial composition, element types, combination patterns, and the proportion of elements. Consequently, the traditional design process is not only intricate but also highly dependent on the designer’s accumulated experience, thus typically resulting in a lengthy design cycle of one to three months or even longer (Figure 12).

However, the generation mechanism of CGAN actuates the inheritance of the complex spatial characteristics of gardens through algorithmic operations. Its visual output not only decodes the spatial topological rules of regional gardens but also illustrates the spatial composition patterns under the influence of regional cultural contexts. By deploying the well-trained CGAN model, designers can spend less energy on studying the characteristics of classical gardens. It only takes 5 to 10 s to generate 30 to 50 different scheme floor plans using the CGAN model trained in this study. This has shortened the design cycle from the traditional 3 months to 10 s. This represents a shift from human-scale deliberation (weeks/months) to machine-speed synthesis (seconds), fundamentally transforming the design paradigm (Figure 13).

CGAN is not only a research medium for the spatial cognition of traditional culture, but also an efficient decision-making tool for scheme design. It can play a key role in shortening the scheme design cycle in historic site protection and renewal.

4.2. Multi-Scheme Comparison of the Planning Site

The model trained in this study supports the generation of multiple types of schemes. Designers or decision-makers can compare and select multiple schemes based on the characteristics and types of the site. This model can significantly shorten the design cycle while ensuring that the scheme maintains coherence with regional traditions in terms of spatial organization logic. With the help of this trained CGAN model, four different types of landscape spaces can be generated according to the characteristics of the spatial environment (rockery sculpture space, water-island space, pavilion gallery space, and botanical spaces).

(1)

Different schemes are generated for comparison and selection at the same site.

The spatial size of the training samples for this model ranges from 200 to 1000 square meters. The floor plans generated by CGAN can be applied in the renewal of micro and small spaces in urban historic protection blocks; for example, micro-space updates (200–500 m²), similar in size to pocket parks, street corners, and tiny green spaces. Small-scale space updates (500–800 m²) are similar to community plazas and courtyards. The medium-scale space update (>800–1000 m²) is similar to neighborhood parks. Decision-makers can select different types of models based on the size of the site space to generate multiple landscape space floor plans on the same site. Thus, decisions and comparisons can be made based on the actual site requirements (Figure 14).

(2)

Different venues generate different scheme comparisons.

Against the complex background of urban renewal and public space planning, decision-makers often face the core predicament of the synchronous and overall planning of multiple sites. They must conduct simultaneous design evaluations for multiple spaces, such as pocket parks, community squares, block green spaces, etc. The traditional method relies on manual sequential analysis, resulting in low efficiency and difficulty in unifying standards. Meanwhile, the design floor plans for different venues are proposed by different teams, with significant differences in style, function, or cost. Therefore, decision-makers lack a standardized comparison framework and find it difficult to quantify the advantages and disadvantages (Figure 15).

With the CGAN, the specific difficulties of multi-site collaborative planning can be solved. The CGAN model trained in this study can automatically generate design schemes according to different spatial types with a unified style and differentiated landscape functions. The CGAN transforms multi-site planning from “empirical manual labor” to “multi-type intelligent generation.” This not only effectively leverages the decision-makers’ experience and efficiency but also intuitively assists in the generation of design decisions through visual floor plans. This technology can provide a scalable new paradigm for the fragmented space renewal of high-density historic and culturally protected areas.

4.3. Evaluation of Space Utilization Rate

The scheme floor plan generated by CGAN marks landscape types in different colors (comprising 12 landscape elements such as architecture, pavilions and structures, pergolas, water areas, trees, etc.). The land-use ratio of each landscape element in the scheme can be calculated through the color proportions. Decision-makers can evaluate the space utilization rate of the generated floor plan by comparing it with the park design specifications based on the actual site type. Finally, the best scheme with relatively rich landscape elements that meets the design specifications is selected.

When multiple schemes are executed simultaneously, decision-makers can first create different spaces according to the size and outline of the site (rockery sculpture space, water-island space, pavilion gallery space) and spatial generation (botanical spaces). Subsequently, all the generated landscape elements and technical indicators are calculated individually. This way, the types and proportions of the landscape elements and the space utilization rate of each generated result can be obtained. Finally, by comparing the space utilization rate with the land-use indicators in the park design specifications, the scheme floor plan that meets the design specifications can be identified (Figure 16).

We take three project sites of different scales as examples. Among the generation results for the micro-space, the pavilion corridor space was the most abundant and conformed to the park design specifications. According to China’s Code for the Design of Public Park (GB 51, 192–2016) [40], for small parks with an area of less than 2 hectares, the proportion of landscape greening (including water areas) should exceed 65%, the proportion of buildings should be less than 1.0%, and the land for garden paths and paving should be between 15% and 30%. The pavilion corridor space conformed best to the specifications (the proportion of landscape greening was 82.61%, the proportion of buildings was 0%, and the proportion of garden paths and paving was 17.39%).

Meanwhile, according to the Park Design Code (2016), for community parks with an area of less than 2 hectares, the proportion of landscape greening (including water body areas) must be > 65%, and the proportion of buildings should be <3% (with the proportion of management buildings less than 0.5%, the proportion of service buildings less than 2.5%, and the land for garden paths and paving between 15% and 30%). Among the four results for the small-scale space, the water-island space agreed most with the norms (the proportion of landscape greening was 68.53%, the proportion of buildings 2.47%, and the proportion of garden paths and paving 29.01%).

Furthermore, among the results for the medium-scale space, the pavilion gallery space was the most abundant. It complied with the park design standards (landscape greening accounted for 71.69%, buildings for 2.91%, and garden paths and paving for 25.37%).

Thus, decision-makers can compare the space utilization rates of the different generated results, refer to the design specifications, and form the best design decision. In practical applications, they can select the appropriate spatial type. Based on the output results of the model and the requirements, they can prioritize the schemes with rich landscape elements and a high space utilization rate and evaluate and adjust the floor plan to meet the needs of urban renewal.

4.4. Satisfaction with the Generated Floor Plan

4.4.1. Investigation Methods and Verification

To further evaluate the satisfaction of designers with respect to the garden space floor plans generated using the CGAN, the researchers selected one group of typical garden spaces from each of the four spatial types, forming a total of four groups of comparison schemes (each group included one original scheme and one CGAN-generated scheme), for a total of eight design schemes as the investigation samples. The satisfaction survey was conducted from five evaluation perspectives: the rationality of the layout, the adaptability of the scale, the richness of the content, the aesthetic quality, and the sense of layering in the landscape (Figure 17).

The survey was scored using a 5-point Likert scale: 5 points = very satisfied/very much in agreement; 4 points = relatively satisfied/relatively in agreement; 3 points = average/acceptable; 2 points = not very satisfied/not quite in agreement; and 1 point = dissatisfied/disagree. Under double-blind testing conditions (without revealing the source of the scheme), designers and decision-makers were invited to independently score the scheme based on their professional judgment and personal preferences (Appendix C). Comparing the score differences between each group of original schemes and the CGAN-generated schemes, the designers’ acceptance of the AI-generated schemes was analyzed.

A total of 310 questionnaires were distributed in this study. After screening and eliminating 4 invalid questionnaires, 306 valid questionnaires were finally obtained. To ensure the reliability and consistency of the research data, we conducted reliability and validity analyses of the questionnaire using the SPSS 21 statistical software. Reliability is used to measure the stability of questionnaire results and the authenticity of the measured features. Cronbach’s alpha coefficient is the most commonly used internal consistency evaluation index, which can effectively reflect the reliability level of the questionnaire [41]. The value of the standardized Cronbach’s alpha coefficient ranges from 0 to 1; the closer the coefficient is to 1, the higher the consistency within the questionnaire and the more reliable the results. It is generally believed that a questionnaire with a standardized Cronbach’s alpha coefficient value greater than 0.6 is sufficiently reliable [42].

The reliability results of this questionnaire are shown in

Table 3. Reliability statistics.

Cronbach’s Alpha	Cronbach’s Alpha Based on Standardized Items	No. of Items
0.924	0.923	40

. The Cronbach’s alpha coefficient of all measurement dimensions was 0.924, which is greater than the standard of 0.6, indicating that the questionnaire had good overall reliability (Table 3).

Furthermore, to ensure the validity of the questionnaire measurement results, the validity of the questionnaire was also tested. Validity reflects the extent to which the measurement tool can accurately assess the characteristics of the target. The validity level was evaluated by calculating the KMO (Kaiser–Meyer–Olkin) value. The value range of this index is 0–1. The closer the value is to 1, the better the structural validity of the questionnaire, and the more truly the measurement results can reflect the characteristics of the research object [43]. In this study, the validity analysis of the questionnaire mainly applied the KMO and Bartlett sphere tests to illustrate whether the internal structure of the questionnaire was valid. The KMO and Bartlett values are considered the best indicators of validity. When the KMO test coefficient is > 0.5 and the Bartlett significance level is <0.05, the questionnaire has structural validity [44].

According to the results in

Table 4. KMO and Bartlett’s test results.

Kaiser–Meyer–Olkin Measure of Sampling Adequacy	Bartlett’s Test of Sphericity
	Approx. Chi-Square	df	Sig.
0.902	5728.149	780	0.000

, the KMO value of this questionnaire was 0.902, Bartlett’s approx. Chi-square was 5728.149, the degree of freedom (df) was 780, and Bartlett’s significance (Sig) was 0.000, which is less than 0.05, indicating that this questionnaire has good authenticity and validity (Table 4).

4.4.2. Analysis of Investigation Results

(1)

Overall satisfaction comparison

According to the statistical results of the total average value of each group of spatial evaluations, the average satisfaction score for all floor plans was 3.871, which is greater than the level of 3 (general). Among them, the average satisfaction scores generated by the CGAN model were mostly higher than the average scores for the original floor plan. In the floor plans generated by CGAN, the average satisfaction score for the water-island space was the highest at 4.080, while the satisfaction score for botanical spaces was relatively low, at 3.659. In the original floor plan, the average satisfaction score for the water-island space was the highest, at 4.006, while for the botanical spaces it was the lowest, at 3.628 (

Table 5. Item statistics.

	Name of Shape Group	Source	Mean	Std. Deviation	No. of Items
1	Rockery sculpture space	Authentic garden space	3.882	1.007	306
		CGAN model’s design	3.712	1.078	306
2	Water-island space	Authentic garden space	4.006	0.969	306
		CGAN model’s design	4.083	0.943	306
3	Pavilion corridor space	Authentic garden space	3.943	0.974	306
		CGAN model’s design	4.063	0.966	306
4	Botanical spaces	Authentic garden space	3.620	1.156	306
		CGAN model’s design	3.659	1.132	306

).

According to the comparison scores for the different type groups, the average scores of the three shape groups in the floor plan generated by the CGAN model were all higher than the scores of the original floor plan. In the rockery sculpture space comparison group, the average satisfaction score of the CGAN-generated planar floor plan was slightly lower than that of the original planar floor plan (3.882) and the CGAN-generated planar floor plan (3.712) (Figure 18).

(2)

Analysis of performance in each dimension

Rationality of Layout: In the CGAN scheme, the water-island space (4.02) and the space around the building (4.00) scored the highest, and the performance in the rock space (3.62) was relatively weak.

Adaptability of Scale: The water-island space (CGAN scheme: 4.20) and the space around the building (CGAN scheme: 4.20) performed the best, indicating that CGAN had more advantages in coordinating the spatial proportion.

Richness of Content: The water-island space (CGAN: 4.14) received the highest score, while the flower and tree space (CGAN: 3.73) and the rock space (CGAN: 3.82) received lower scores.

Aesthetic Quality: Among all types, the water-island space (CGAN scheme: 3.96) and the space around the building (CGAN scheme: 3.92) had relatively high aesthetic evaluations, but the flower and tree space (CGAN scheme: 3.49) and the rock space (CGAN scheme: 3.52) performed relatively weakly.

Sense of Layering: The CGAN scheme performed well in terms of sense of layering in the water-island space (4.09) and the space around the building (4.07), but needs to be strengthened in the rock space (3.72) and in the design of spatial depth (Figure 19).

The comprehensive analysis indicated that the garden floor plan schemes generated by CGAN performed exceptionally well in terms of overall satisfaction, with an average score (3.879) higher than the authentic garden space (3.863).

Among them, the CGAN design schemes for the water-island space and the pavilion corridor space had the highest satisfaction rates (4.083 and 4.063, respectively). In dimensions such as scale coordination (4.20), content richness (4.14), and layering (4.09), CGAN has prominent advantages, demonstrating the potential of AI in optimizing spatial proportion and landscape structure. However, satisfaction with the CGAN scheme for the rockery sculpture space (3.712) was slightly lower than that for the original design (3.882), and it performed poorly in terms of layout rationality and aesthetic appeal. This indicates that simulating complex natural forms with the algorithm needs to be improved. Furthermore, the overall evaluation of the botanical spaces was relatively low (CGAN scheme: 3.659), reflecting the limitations of the existing models in terms of capturing the content richness and visual aesthetics of the botanical spaces.

Future research will focus on optimizing the layout logic of rock and mountain spaces and enhancing the content richness of flower and tree spaces. At the same time, it will maximize CGAN’s technical advantages in water areas and the spaces around buildings to comprehensively promote the application of AI-assisted design in landscape planning.

5. Conclusions

This research studied the Suzhou Gardens, a World Cultural Heritage Site, with the aim of exploring innovative methods using artificial intelligence technology to assist in decision-making regarding the design of garden spaces. It provides efficient and scientific tools for the protection and renewal of traditional gardens by urban decision-makers restricted by the limitations of current garden design, such as reliance on manual experience, low efficiency, and insufficient continuity of traditional styles. This study proposes an intelligent generation framework based on CGAN to enhance the feasibility and innovativeness of garden space design decisions.

The main innovations of this research are reflected in the following three aspects: Firstly, we analyzed the core characteristics of traditional garden spaces. According to these characteristics, the four main spatial types that constitute Suzhou Gardens were systematically summarized, including rockery sculpture space, water-island space, pavilion gallery space, and botanical spaces. Meanwhile, color annotations were applied to the 12 main landscape space elements. Secondly, a deep learning architecture based on CGAN was designed. This model adopted the adversarial training mechanism, including a generator and a discriminator. In terms of the training strategy, the “contour atlas” was taken as the conditional sample and the “spatial element layout atlas” as the generation sample to train the layout generation of spatial elements under the control of contour conditions. By observing the loss logs (loss values) during the training process and the test images in the iteration, we ensured that the generated results not only conformed to the data distribution but also maintained spatial rationality. Finally, we applied the proposed model to generate design decisions. Application methods such as the scheme design process assisted by machine learning, the multi-scheme comparison and selection mode, and the space utilization rate evaluation were proposed, providing designers with an efficient and multidimensional design decision-making basis.

The main conclusions of this study are as follows:

(1)

The CGAN model can effectively capture the key layout characteristics of the garden space. Four different types of landscape spaces, namely, rockery sculpture space, water-island space, pavilion gallery space, and botanical spaces, can be generated according to the characteristics of the spatial environment.

(2)

This model can be applied to projects with space sizes ranging from 200 to 1000 square meters. The plans generated by CGAN can be applied in the renewal of micro and small spaces in urban historic conservation blocks.

(3)

After 500 epochs of training, the CGAN model can accurately learn the organizational rules of garden spaces, and the average accuracy rate of the generated planar schemes reached 91.08%.

(4)

The efficiency of the scheme generation was significantly improved. Compared with traditional design, which relies on experience and takes 1 to 3 months, 30 to 50 floor plans can be generated within 3 to 10 s on the same site. This has greatly improved the design efficiency.

(5)

CGAN can provide multiple scheme comparisons for decision-making in urban historic space renewal, helping managers to select the optimal solution in a short period of time.

(6)

The land-use proportions of 12 spatial elements can be quickly calculated based on the color of the element annotations. From this, the design specifications can be referred to for a quantitative assessment of the generated space utilization rate. Among the generated floor plans, the proportion of landscape greening basically met the specification requirements (>65%).

(7)

The floor plan generated by CGAN was slightly higher than the original floor plan in terms of the overall average satisfaction. The average satisfaction with the CGAN scheme (3.879) was higher than with the authentic garden space (3.863).

This study can help decision-makers efficiently generate and evaluate the feasibility of different design schemes, providing intelligent support for decision-making in urban renewal plans. However, this model remains limited in depicting small-scale details, and the data and algorithms need to be further optimized. The following deficiencies remain:

(1)

The detailed processing ability of the model needs to be improved to generate clear and realistic image results.

(2)

The model still faces certain challenges when dealing with the planar contours of curved surfaces and requires further improvement and optimization.

(3)

There is room for improvement in the clarity and cleanliness of the generated images.

(4)

The layout logic of building and rockery spaces must be optimized, the content richness of botanical spaces enhanced, and CGAN’s advantages in water areas consolidated.

In summary, generative adversarial networks (GANs) have shown great potential in the field of landscape design decision-making due to their powerful image generation capabilities. The significance of this research relates to providing an intelligent technical path for the protective development of historic gardens. In particular, the digital inheritance of traditional construction wisdom can be achieved through machine learning methods. On the other hand, the decision support system constructed in this study can assist managers in quickly evaluating the feasibility of different strategies in the early stage of scheme design. It can effectively balance the contradiction between protection and development.

In the future, researchers will further improve and optimize the CGAN model to enhance its robustness and stability. Existing planar generation results can also be combined with 3D model generation tools to generate 3D model spaces, offering the functional expansion from floor plan generation to 3D model generation. We can explore methods of collaborative design decision-making between models and designers to further improve the design quality and achieve better human–computer interactions. This technology can be further extended to the field of protection and renewal of historic districts in other cities, providing intelligent support for urban planning decisions.