1. Introduction
Urban thermal environments and the urban heat island (UHI) effects have emerged as critical challenges affecting both the quality of life and the sustainable development of cities. The intricate interplay between urban morphology and local climate conditions significantly influences temperature distribution, energy efficiency, and the intensity of UHI phenomena [1]. Consequently, the generation of high-quality 3D urban morphology models represents an essential precursor for comprehending and simulating urban thermal dynamics. Zhang et al. [2] employed Landsat 8 imagery alongside a 3D compactness index to elucidate the effect of 3D urban compactness—comprising building density, height, and their synergistic interactions—on the urban heat environment. Their findings informed the development of an optimized urban morphology framework aimed at enhancing temperature regulation and spatial thermal distribution. Deng et al. [3] investigated the thermal performance of three archetypal urban configurations and associated morphologies, revealing that dispersed urban forms exhibited superior outdoor thermal conditions. Similarly, Ramyar et al. [4] analyzed a suite of urban cooling strategies and heat mitigation interventions, proposing adaptive design solutions and morphology-oriented strategies tailored to the unique demands of Tehran’s dense urban structure.
Conventional methods for investigating urban thermal environments often rely on preexisting imagery coupled with labor-intensive manual analyses. These approaches present significant challenges, particularly in generating urban morphologies that correspond to diverse urban forms and thermal scenarios with efficiency and precision. Recent advancements in image processing and generative technologies have revolutionized this domain, enabling the rapid and systematic creation of urban morphologies. Zhou et al. [5] demonstrated the efficacy of Generative Adversarial Networks (GANs) in producing diverse urban configurations tailored to specific local climate conditions, thereby markedly enhancing the efficiency, scalability, and adaptability of data generation and modeling processes. Traditional methods, however, have been constrained by their focus on localized urban clusters or small-scale analyses, rendering their findings less transferable to broader regional contexts or multi-scenario modeling. Modern computational advancements, including high-performance hardware and sophisticated software platforms, have effectively mitigated these limitations. Furthermore, by expanding the scope and diversity of GAN training datasets, it is now feasible to construct urban morphological models with versatility to address a wide spectrum of urban and climate scenarios, significantly broadening their practical applicability [6,7].
Among the diverse GAN architectures, Pix2Pix [8,9] stands out as a conditional GAN designed for image-to-image translation, offering distinct advantages in urban morphology generation and simulation [10,11]. The Pix2Pix framework comprises a generator that transforms input images and a discriminator that evaluates and distinguishes the disparities between the generated outputs and actual target images [12,13]. While Pix2Pix demonstrates notable efficiency in performing conditional image-to-image translation, it occasionally struggles to encapsulate the complexity and granularity inherent in urban forms and layouts [14,15]. This limitation often manifests as outputs with diminished realism and visual coherence [16].
To address these shortcomings, the integration of Pix2Pix with CycleGAN presents a robust and innovative solution [17]. CycleGAN [15], originally conceived for image style transfer tasks, excels in unsupervised style transformation, enabling conversions across differing image styles without necessitating paired datasets [18]. By leveraging the outputs generated by Pix2Pix alongside actual images as dataset pairs, CycleGAN facilitates the refinement of urban morphology representations, enhancing realism and fidelity [19,20]. A key feature of CycleGAN is its cycle-consistency loss, which ensures that transformed images retain a high degree of correspondence to their real-world counterparts [21]. The synergy between Pix2Pix and CycleGAN significantly improves the quality and precision of generated urban morphologies. Pix2Pix effectively establishes the foundational structure, while CycleGAN refines the outputs by capturing intricate details and enhancing visual authenticity [21,22,23]. This coupled approach offers a powerful pathway for generating highly realistic urban environments, enabling more accurate simulations and analyses of complex urban forms.
Urban morphology lacks embedded climatic data, necessitating the integration of local climatological characteristics for comprehensive modeling. Stewart and Oke [24] introduced the local climate zone (LCZ) classification scheme as a systematic framework for urban climatologists to describe and characterize the physical attributes of cities. The LCZ scheme comprises 17 zones, defined primarily by surface structure properties (e.g., building and vegetation height and density) and surface cover types (e.g., pervious versus impervious surfaces) [24,25]. Each zone is local in scale, offering a standardized and logical starting point for assembling consistent and comparable urban climate data across diverse urban clusters [26]. For instance, LCZ1 corresponds to high-density urban areas, while LCZ5 represents suburban areas with extensive vegetation coverage. Integrating LCZ data into GAN-based urban morphology modeling significantly enhances the capability to generate urban forms tailored to specific climatic contexts [27]. This integration not only facilitates the generation of physical urban forms but also enables the customization of urban landscapes in alignment with predefined climate zones [28]. By doing so, the resulting models achieve greater relevance and applicability, particularly for real-world urban planning initiatives and thermal environment analyses [29]. Such advancements underscore the transformative potential of combining urban climatology with generative modeling techniques to address the multifaceted challenges of urban thermal environments and climate-sensitive urban design [30].
This paper introduces an interface designed to enhance accessibility and usability for a suite of well-established software tools. The interface integrates advanced visualization capabilities and streamlined data input processes, facilitating intuitive interactions for users. By incorporating the synergy between Pix2Pix and CycleGAN models, this interface enhances the generation of high-fidelity urban morphology simulations, tailored to specific local climate zones (LCZs). To demonstrate its practical applicability, this study incorporates real-world case studies and simulation results, showcasing the interface’s effectiveness in addressing complex urban challenges. This integration significantly improves the adaptability and precision of urban planning and thermal environment management. By enabling urban planners to efficiently comprehend and deploy diverse analytical tools, this interface serves as a powerful platform for optimizing and managing urban thermal environments with precision and agility [31].
2. Materials and Methods
The proposed methodology integrates Pix2Pix and CycleGAN architectures, leveraging advanced data-driven modeling techniques to analyze and predict spatial patterns of urban morphology corresponding to the specified LCZs. The methodology is structured into four key sections (Figure 1):
-
(1). Study Area and Data Preparation
This study focused on eight cities in China—Dalian, Tianjin, Qingdao, Ningbo, Xiamen, Shenzhen, Zhuhai, and Haikou—selected for their diverse urban morphologies and climatic conditions. Data collection involved compiling LCZ classification maps and the corresponding urban morphology dataset, which were preprocessed to ensure consistency and suitability for subsequent modeling tasks.
-
(2). Data Augmentation and Dataset Construction
To enrich the dataset and enhance model training efficacy, data augmentation techniques were employed, resulting in a robust dataset comprising 4712 samples. This dataset was partitioned into 80% for training and 20% for testing, ensuring a balanced approach to model development and evaluation while mitigating overfitting and underfitting risks.
-
(3). Model Training and Result Generation
The core of the methodology lies in the synergistic application of Pix2Pix and CycleGAN neural network architectures. These models were trained to establish mappings between input data (e.g., LCZ maps and urban morphology characteristics) and target outputs, capturing the nuanced relationships within the data. Model performance was rigorously evaluated using quantitative metrics, including the Structural Similarity Index Measure (SSIM) and the R2score [32]. Iterative training processes were implemented, wherein models were retrained as needed until satisfactory performance metrics were achieved, ensuring high-quality outputs [33].
-
(4). 3D Model Reconstruction
To translate the generated 2D urban morphology data into three-dimensional representations, the Grasshopper platform within Rhino 7.9 was utilized. This step involved the reconstruction of high-precision 3D meshes, providing a detailed and visually accurate depiction of urban areas, thus facilitating a more immersive analysis and simulation of urban morphology and its interaction with local climate conditions.
2.1. Data Acquisition
The dataset utilized in this study encompasses representative cities across China, including Dalian, Qingdao, Tianjin, Ningbo, Xiamen, Shenzhen, Zhuhai, and Haikou. A geographical overview of the selected cities is provided in Figure 2. Urban form data were sourced from Mapbox Streets v7 (
The World Urban Database and Access Portal Tools (WUDAPT) initiative is closely aligned with the LCZ framework, serving as a global platform for the collection, analysis, and dissemination of urban climate data related to physical geography [34]. In this study, Level 0 LCZ classification data from WUDAPT were employed for eight selected cities, providing a foundational dataset for the experimental analyses. The accuracy of the LCZ classifications was evaluated through a comprehensive set of metrics, including overall accuracy (OA), urban LCZ accuracy (OAu), built versus natural LCZ accuracy (OAbnat), and weighted accuracy (OAw). These metrics offer a rigorous framework for assessing the reliability and validity of the classification data across a variety of urban contexts [35,36]. The LCZ classification data with the highest accuracy in each city was downloaded from the WUDAPT Level 0 platform and detailed accuracy assessments for each city are presented in Table 1 and Figure 3.
2.2. Data Preprocessing
The LCZ classification map and urban morphology map were exported from ArcGIS as 256 × 256 pixel tiles with a resolution of 300 DPI. The LCZ map was designated as Dataset A, while the urban morphology map was labeled as Dataset B. These datasets were merged in a left–right configuration, with Dataset A on the left and Dataset B on the right, to create a combined dataset suitable for Pix2Pix processing. For the CycleGAN dataset, the tiles generated by the trained Pix2Pix model, along with the original tiles, were utilized as training data. To enhance model generalization and robustness against variations in input data, data augmentation techniques were applied [37]. For Pix2Pix, these techniques included 90° rotations, 270° rotations, and mirroring [38]. For CycleGAN, random mirroring and partial scaling were employed to increase the diversity of training samples. These augmentation strategies aim to reduce the risk of overfitting by exposing the models to a wider range of input variations [39], as illustrated in Figure 4. This preprocessing pipeline ensures the creation of a robust and diverse dataset, facilitating the effective training and optimization of the Pix2Pix and CycleGAN models.
2.3. Structure of Generative Adversarial Network
GANs, introduced by Goodfellow et al., represent a cutting-edge advancement in deep learning. The GAN framework consists of two primary components: the generator (G) and the discriminator (D). These networks operate in an adversarial manner, wherein the generator seeks to produce samples that closely resemble real-world data, while the discriminator evaluates the authenticity of the generated samples and assigns a probability score [40]. This iterative adversarial process continues until the generator produces samples of such high quality that the discriminator can no longer reliably differentiate between real and generated data. The generator (G) transforms a noise vector into new samples using operations such as convolution, deconvolution, and fully connected layers. Simultaneously, the discriminator (D) processes the input samples, assesses their authenticity, and outputs a probability value indicating the likelihood that the sample is real. This dynamic interaction drives the optimization of the network parameters and the progressive refinement of the generated outputs [41].
Among the advancements in GAN technology, the Pix2Pix and CycleGAN algorithms represent significant milestones. The combined workflow of Pix2Pix and CycleGAN offers multiple advantages across four key aspects [42]. First, this integration combines the strengths of supervised and unsupervised learning. Pix2Pix generates structurally precise images using paired data, while CycleGAN performs style transformations without requiring paired datasets, thereby enhancing the realism of the generated images. This adaptability allows the workflow to accommodate varying data requirements effectively [43]. Second, the sequential application of Pix2Pix and CycleGAN improves image quality [44]. Pix2Pix is employed to create foundational structural images, which are then refined by CycleGAN to enhance stylistic details. This stepwise approach significantly elevates the overall quality of the generated outputs [45]. Third, the use of data augmentation techniques in both stages enhances the diversity of generated samples and strengthens model robustness, reducing susceptibility to overfitting. Finally, the combination achieves a balance between structural accuracy and stylistic authenticity. Pix2Pix ensures that the image structure remains precise, while CycleGAN enriches the visual appeal without compromising structural integrity. This synergy enhances the quality of generated images, making the method highly effective across diverse application scenarios [46].
The workflow illustrated in Figure 5 demonstrates the Pix2Pix process for structural shaping followed by CycleGAN’s style enhancement, showcasing the method’s ability to produce realistic and high-quality outputs. This integrated approach highlights its utility in generating visually compelling and structurally accurate representations in various domains.
2.3.1. Training Framework of Pix2Pix
Pix2Pix represents a significant innovation introduced by Isa et al. [8] in the domain of image-to-image translation. Its architecture is based on a GAN framework, comprising a generator (G) and a discriminator (D). The generator (G) employs the U-Net architecture, which is particularly well suited for image processing tasks due to its encoder–decoder structure and skip connections that facilitate precise feature mapping between input and output images [47].
-
(1). Architecture and Functionality
The U-Net architecture within Pix2Pix consists of two primary components, the encoder and the decoder, interconnected by skip connections. The encoder progressively extracts high-level semantic features from the input image while simultaneously reducing its spatial resolution. This process is achieved through the sequential application of convolutional layers, batch normalization, and the Rectified Linear Unit (ReLU) activation function. These operations allow the encoder to capture essential features required for accurate image translation. The decoder, in turn, reconstructs the extracted features into the target image. It employs deconvolution layers, with normalization applied to scale outputs to the range [−1, 1]. Additionally, the decoder incorporates regularization techniques, such as the Tanh activation function and Dropout, to improve the reconstruction accuracy of the output image and prevent overfitting [48]. The integration of skip connections enhances the decoder’s ability to preserve fine-grained details from the input image, further improving the quality of the generated outputs [49].
-
(2). Loss Functions and Training Process
The training phase of Pix2Pix places significant emphasis on monitoring the loss curves of both the generator and discriminator to ensure optimal performance [50]. Key loss components include Adversarial Loss and L1 Loss. Adversarial Loss: This component ensures that the generator produces images indistinguishable from real images by deceiving the discriminator. The discriminator evaluates the authenticity of generated images and provides feedback to the generator for refinement [51]. L1 Loss: This loss minimizes the pixel-wise differences between the generated image and the ground truth, promoting accurate structural reconstruction in the output. Gen_total Loss, representing the overall loss of the generator, is also tracked to evaluate model convergence [52].
-
(3). Objective and Impact
The primary objective of the generator in Pix2Pix is to produce images so realistic that the discriminator cannot reliably differentiate between real and generated images. This adversarial dynamic drives the iterative refinement of both networks, culminating in high-quality image translation [53]. By combining the adversarial loss to enhance realism and the L1 loss to ensure structural accuracy, Pix2Pix delivers robust and visually consistent outputs, establishing itself as a foundational method in image-to-image translation applications [54].
2.3.2. Training Framework of CycleGAN
The CycleGAN architecture was constructed with two generators (GA and GB) and two discriminators (DA and DB). As an unsupervised learning model, CycleGAN facilitates transformations between two domains, A and B, without the need for paired training data [55]. The cornerstone of this framework is the cyclic consistency loss, which ensures that information is preserved during bidirectional transformations. This loss guarantees that an image transformed from domain A to B and then back to A retains its original characteristics [56].
The predicted image styles generated by Pix2Pix were refined using CycleGAN, aligning them more closely with the realistic urban morphology styles. Simultaneously, the discriminator (D) assessed the similarity between the generated images and the real images, providing feedback to the generators for iterative refinement.
-
(1). Loss Functions in CycleGAN
The CycleGAN framework incorporates two primary types of loss functions. Adversarial Loss: This component drives the competition between the generators and discriminators. The goal of the generators (GAB and GBA) is to produce images that are indistinguishable from real images within the target domains, while the discriminators (DA and DB) strive to accurately differentiate real images from generated ones. Cyclic Consistency Loss: This loss ensures that after a round-trip transformation (e.g., A→B→A), the reconstructed image remains consistent with the original. This mechanism preserves critical structural and stylistic information, enhancing the fidelity of the transformation process.
-
(2). Training Dynamics
During training, the discriminators (DA and DB) continuously learn to identify real images and distinguish them from generated ones. Correspondingly, the generators (GAB and GBA) are optimized to create realistic images that deceive the discriminators, effectively narrowing the gap between the real and generated distributions. This adversarial dynamic underpins the iterative improvement in both generators and discriminators, ultimately producing high-quality, realistic transformations.
The CycleGAN training framework, with its integration of adversarial and cyclic consistency losses, enables unsupervised domain adaptation while preserving critical image features. This methodology proves particularly effective for style adjustments and refinement of urban morphology images, as demonstrated in the context of this study [57].
2.3.3. The Hyperparameter Settings of the Pix2Pix and CycleGAN Algorithms
The hyperparameter settings play a crucial role in training Pix2Pix and CycleGAN models as they directly affect the model’s stability, convergence speed, and the quality of generated images. Batch size determines the number of samples used in each iteration; smaller batches allow for faster updates but may lead to instability, while larger batches provide smoother training [58]. The learning rate controls the step size for weight updates, and an appropriate learning rate helps accelerate convergence while avoiding instability. Weight decay, as a regularization technique, helps prevent overfitting and ensures that both the generator and discriminator generalize well [59]. Lambda parameters balance the contributions of different loss components in the objective function, ensuring high-quality and consistent image generation [60]. Beta parameters adjust the gradient momentum in the Adam optimizer, helping to smooth the training process and avoid gradient explosion or vanishing. Ema_momentum smooths model weight updates, reducing noise and promoting the generation of more realistic images. Finally, Total epochs and best epochs control the total number of training rounds and the saving of the model at its optimal performance. For the specific hyperparameter settings used in this study, please refer to Table 2. Proper tuning of these hyperparameters is essential for improving the model’s training efficiency and generating high-quality, stable images.
2.3.4. Accuracy Assessment
To effectively evaluate and compare the generalization ability of the models, accuracy assessments were conducted using a consistent test set for both algorithms. The evaluation framework incorporated two dimensions: quantitative metrics and qualitative analysis. For quantitative evaluation, the Structural Similarity Index (SSIM) and the coefficient of determination R2 were employed as primary metrics [61]. The SSIM measures the perceptual similarity between the generated and ground truth images, capturing structural information and visual coherence. The R2 score evaluates the statistical correlation between the predicted and actual data, providing insights into the model’s predictive accuracy. For qualitative evaluation, the authenticity and fidelity of the generated images were assessed through visual inspection. This approach involved a detailed comparison of the generated outputs against ground truth images to subjectively evaluate their realism and alignment with actual urban morphology styles [62].
3. Results
3.1. Supervision and Summary of the Training Process
Figure 6 presents the loss curves for the Pix2Pix model, illustrating its training dynamics. The Pix2Pix model underwent a total of 1000 training iterations. The loss trajectories of the generator (G) and the discriminator (D) serve as quantitative indicators of the model’s performance throughout the training process. During the training of the Pix2Pix model, the loss curves exhibited a general downward trend, converging over time.
In the initial training stages, the discriminator’s loss curve displayed substantial fluctuations; however, as training progressed, it gradually stabilized and ultimately converged at a lower loss value (~0.5). This stabilization suggests that the discriminator incrementally refined its ability to differentiate between real and generated images, thereby reaching a state of convergence [62]. The generator’s loss curve remained relatively stable, reaching a minimum loss of 0.19, which indicates continuous optimization in image generation. As the loss progressively declined, the generated images increasingly resembled real images, making it increasingly difficult for the discriminator to distinguish between them—a fundamental objective of adversarial training [63].
Key metrics, such as Total GAN Loss and Generator L1 Loss, were employed to assess the quality of the generated images. A lower Total GAN Loss signifies that the generated images exhibit greater fidelity to real images, rendering them more challenging for the discriminator to distinguish. As depicted in Figure 6, both Total GAN Loss and Generator L1 Loss stabilized after approximately 150 iterations, underscoring the progressive convergence of the model. By the 811th iteration, these losses reached their respective minima—Total GAN Loss at 0.19 and Generator L1 Loss at 0.10—indicating a significant enhancement in the model’s adversarial training efficacy [64].
For the CycleGAN model, the SSIM and R2 were selected as evaluation metrics, as the primary objective of image translation tasks extends beyond the mere optimization of loss values to encompass fidelity and structural preservation of the generated images [65]. The CycleGAN loss function comprises multiple components—including generator loss, discriminator loss, and cycle-consistency loss—each of which carries different weightings during model training. Consequently, numerical variations in the loss function may be influenced by these weight adjustments and thus may not always serve as a definitive measure of image quality. In contrast, SSIM and R2 are designed specifically to quantify content fidelity and structural coherence, making them more appropriate for assessing the model’s performance in this context [63]. As shown in Figure 7, CycleGAN achieved a peak SSIM of 0.918 and an R2 of 0.987 at the 340th iteration on the test set, demonstrating high structural similarity and predictive accuracy.
For experimental validation of the Pix2Pix-CycleGAN model, the checkpoint from the 811th iteration was initially designated as the benchmark. Given that Pix2Pix model checkpoints were saved at intervals of 25 iterations, the model from the 800th iteration was selected for subsequent evaluations. For the CycleGAN model, the checkpoint saved at the 349th iteration was employed for further fine-tuning and experimental analysis. These selections were based on the models’ demonstrated convergence and optimal performance metrics, ensuring robust experimental outcomes.
3.2. Image Evaluation
3.2.1. Analysis of Urban Block Morphology Image Generation
To comprehensively assess the efficacy of CycleGAN-enhanced image generation in comparison to direct Pix2Pix outputs, a random sample was selected from the training set for each coastal city for visual analysis. As illustrated in Figure 8, images generated solely through Pix2Pix training exhibit notable deficiencies in handling edge artifacts. In contrast, images refined through CycleGAN demonstrate superior edge preservation, particularly in delineating building contours and green space boundaries with high fidelity and realism.
A relative error map was employed to quantify the discrepancies between the predicted images and the ground truth, with RGB deviations visualized through a color gradient representing pixel-wise errors. The analysis indicates that CycleGAN-modified images exhibit significantly lower error magnitudes compared to those generated exclusively by Pix2Pix, underscoring the enhanced precision of CycleGAN in image synthesis.
A comparative visualization between the actual and generated images (Figure 8) further elucidates the performance differences between the two models. The second and fourth columns display images produced by Pix2Pix and CycleGAN, respectively, while the third and fifth columns present the corresponding relative error maps. At a macroscopic level, both models effectively reconstruct the overall morphology of buildings and urban blocks, maintaining consistency in outline and color representation. However, upon closer inspection of fine-scale details—particularly along building edges and small-area morphological features—CycleGAN-generated images exhibit sharper contours and smoother edge transitions. Conversely, Pix2Pix-generated images contain localized distortions and blurring in certain regions, indicating limitations in preserving intricate details. In the first, second, and fourth rows, although the Pix2Pix model performs well in image generation, it is still limited by its network architecture, particularly when handling complex and densely detailed data. For instance, in densely packed low-rise building clusters (LCZ3), Pix2Pix may struggle to capture certain details, leading to pixel clustering. On the other hand, CycleGAN-enhanced images can effectively separate these dense low-rise buildings, making the generated urban morphology more realistic and accurate. In the third and fifth rows, compared to the direct output from Pix2Pix, CycleGAN-enhanced images exhibit street orientations and layouts that are much closer to the ground truth, with the urban building orientations aligning more accurately with the real-world data.
The relative error maps further corroborate these findings. The error distribution in CycleGAN-generated images is more localized and exhibits a narrower range, reflecting its enhanced capability in capturing the global structural coherence of urban morphology. In contrast, Pix2Pix-generated images display more dispersed errors, with substantial deviations concentrated along building edges. This suggests that Pix2Pix encounters difficulties in processing complex edge transitions, resulting in artifacts and inconsistencies in high-gradient regions.
3.2.2. Quantitative Comparison of Pix2Pix and CycleGAN Post Modification
SSIM and R2 were employed as quantitative metrics to assess the fidelity and predictive accuracy of the generated images. SSIM evaluates the structural similarity between the generated and reference images, reflecting visual coherence, while R2 quantifies the goodness-of-fit of the regression model, providing an indication of the model’s predictive precision.
As shown in Table 3, CycleGAN-enhanced images show a slight improvement in SSIM and R2 compared to Pix2Pix-generated images, except for R2 in the Xiamen dataset. Although CycleGAN enhances visual details, such as sharper edges and more defined features, these improvements do not significantly affect SSIM and R2. These metrics are more sensitive to global structural similarity than to localized enhancements, so the changes in visual details do not lead to notable improvements in the overall structural composition of the images.
Variations in R2 and SSIM values across different cities can be attributed to the inherent complexity and structural heterogeneity of each urban dataset. This underscores the dependence of the model’s generative stability on dataset-specific characteristics when applied to diverse urban morphologies. The findings reinforce the robustness of the CycleGAN model in producing structurally coherent and visually realistic urban morphology images, demonstrating its advantages over Pix2Pix in most scenarios.
3.3. Model Training and Computational Efficiency
The equipment used for training the model and the training time: The experimental setup included a computer equipped with an Intel(R) Core(TM) i9-9700HQ CPU (Intel, Santa Clara, CA, USA) and an RTX4090Ti GPU (NVIDIA, Santa Clara, CA, USA). We assessed the time required for model loading, preprocessing, and prediction in both the Pix2Pix and subsequent CycleGAN models. In this study, 4539 samples were simulated in total; the Pix2Pix model took approximately 26.4 h, with an average of about 3.3 h per city, while the CycleGAN model took around 23.3 h, averaging 2.9 h per city. The combined model took a total of 49.7 h. Compared to traditional 3D modeling software or machine learning algorithms, the combination of Pix2Pix and CycleGAN demonstrated significant advantages in both computational power and image quality. For urban planners and decision makers, being able to quickly and efficiently simulate and predict urban models based on predefined LCZs will significantly reduce design costs. This method also offers greater benefits for sustainable development.
3.4. Limitations
The integration of CycleGAN to refine the urban morphology model generated solely by Pix2Pix has effectively improved the efficiency of the city morphology workflow based on LCZ input, providing policymakers and urban designers with a lower-cost blueprint for future city planning. Future research should focus on addressing the existing issues within the model to further optimize its performance and enhance its practical applicability.
This study introduced a CGAN by combining Pix2Pix and CycleGAN, thereby improving the accuracy of the original model. However, compared to Pix2Pix, CycleGAN does not require paired data and is trained using unpaired data, which increases the complexity and instability of the training process between the generator and discriminator. Additionally, the use of cycle consistency loss to maintain consistency between input and output images requires the generator to learn more information from unpaired data, further increasing the difficulty and instability of training. With the development of larger models, alternative techniques and more advanced algorithms have emerged, demonstrating significant potential and gaining traction in areas such as 3D building generation and urban planning. These include Variational Autoencoders (VAEs), diffusion models, autoregressive models, normalizing flows, transformer-based architectures, and Energy-Based Models (EBMs).
4. Conclusions
This study explored an advanced approach to urban morphology generation by integrating the Pix2Pix and CycleGAN architectures to enhance the structural fidelity and realism of generated urban forms. Addressing the limitations of conventional image-to-image translation models, the proposed framework successfully leveraged the complementary strengths of both models—Pix2Pix for foundational structural generation and CycleGAN for refinement and edge enhancement. The results demonstrated that CycleGAN-enhanced images exhibited significantly lower relative errors and improved edge delineation, particularly in handling complex urban features such as building boundaries and green spaces. Moreover, by incorporating the LCZ classification framework into the generative modeling process, this study demonstrated the feasibility of producing urban morphologies aligned with climatological conditions. This integration represents a significant advancement in climate-responsive urban modeling, facilitating the generation of more context-aware urban forms.
Conceptualization, M.W. and Z.X.; Methodology, Z.X., S.Z. and Y.W.; Software, Z.X., J.Z. and R.M.A.I.; Validation, Z.X., J.Z. and Y.W.; Formal analysis, M.W. and S.Z.; Investigation, Z.X., J.Z. and L.W.; Resources, M.W., S.Z. and Y.W.; Data curation, Z.X., J.Z. and L.W.; Writing—original draft, Z.X. and L.W.; Writing—review & editing, M.W. and S.K.T.; Visualization, Z.X., J.Z. and L.W.; Supervision, M.W. and S.K.T.; Project administration, M.W.; Funding acquisition, M.W. and L.W. All authors have read and agreed to the published version of the manuscript.
The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding authors.
The authors declare no conflicts of interest.
| Full Form | Abbreviation |
| CycleGAN | Cycle Generative Adversarial Network |
| EBMs | Energy-Based Models |
| GANs | Generative Adversarial Networks |
| LCZs | Local climate zones |
| OA | Overall accuracy |
| OAu | Urban LCZ accuracy |
| OAbnat | Built versus Natural LCZ accuracy |
| OAw | Weighted accuracy |
| Pix2pix | Pixel-to-Pixel |
| ReLU | Rectified Linear Unit |
| SSIM | Structural Similarity Index Measure |
| UHI | Urban heat island |
| WUDAPT | World Urban Database and Access Portal Tools |
| R2 | Coefficient of determination |
| VAEs | Variational Autoencoders |
Footnotes
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Corresponding overall accuracies of eight coastal cities.
| OA | OAu | OAbnat | OAw | |
|---|---|---|---|---|
| Dalian | 0.79 | 0.63 | 0.98 | 0.95 |
| Qingdao | 0.81 | 0.65 | 0.99 | 0.96 |
| Tianjin | 0.78 | 0.77 | 1.00 | 0.96 |
| Ningbo | 0.69 | 0.64 | 0.95 | 0.91 |
| Xiamen | 0.76 | 0.61 | 0.98 | 0.95 |
| Shenzhen | 0.77 | 0.61 | 0.94 | 0.94 |
| Zhuhai | 0.76 | 0.54 | 0.98 | 0.94 |
| Haikou | 0.82 | 0.81 | 0.92 | 0.94 |
Summary of the hyperparameters used for the Pix2Pix and CycleGAN models during the training process.
| Model | Hyperparameters | Value |
|---|---|---|
| Pip2pix | Batch size | 1 |
| Weight_decay_G | 1 × 10−5 | |
| Weight_decay_D | 7.5 × 10−5 | |
| Lambda 1 | 50 | |
| Lambda 2 | 100 | |
| Lambda 3 | 150 | |
| Learning rate_G | 7.5 × 10−5 | |
| Learning rate_D | 4 × 10−5 | |
| Beta_1 | 0.5 | |
| Beta_1 | 0.999 | |
| Ema_momentum | 0.9 | |
| Total epochs | 1000 | |
| Best epochs | 620 | |
| CycleGANs | Batch size | 1 |
| Input size | 256 | |
| Resize scale | 257 | |
| Crop size | 256 | |
| Total epochs | 350 | |
| Decay epochs | 250 | |
| Learning rate_G | 1 × 10−4 | |
| Learning rate_D | 1 × 10−4 | |
| Beta_1 | 0.5 | |
| Beta_2 | 0.999 | |
| Lambda 1 | 10 | |
| Lambda 2 | 10 |
Pix2pix accuracy of eight coastal cities and accuracy after CycleGAN modification.
| Pix2pix | Pix2pix and CycleGAN | |||
|---|---|---|---|---|
| R2 | SSIM | R2 | SSIM | |
| Average | 0.831 | 0.742 | 0.834 | 0.754 |
| Dalian | 0.819 | 0.712 | 0.819 | 0.725 |
| Qingdao | 0.835 | 0.749 | 0.836 | 0.762 |
| Tianjin | 0.819 | 0.719 | 0.820 | 0.731 |
| Ningbo | 0.811 | 0.729 | 0.818 | 0.739 |
| Xiamen | 0.846 | 0.775 | 0.842 | 0.788 |
| Shenzhen | 0.829 | 0.731 | 0.835 | 0.740 |
| Zhuhai | 0.848 | 0.773 | 0.854 | 0.782 |
| Haikou | 0.841 | 0.752 | 0.846 | 0.761 |
References
1. Mobaraki, A.; Oktay Vehbi, B. A conceptual model for assessing the relationship between urban morphology and sustainable urban form. Sustainability; 2022; 14, 2884. [DOI: https://dx.doi.org/10.3390/su14052884]
2. Zhang, X.; Wang, D.; Hao, H.; Zhang, F.; Hu, Y. Effects of land use/cover changes and urban forest configuration on urban heat islands in a loess hilly region: Case study based on Yan’an City, China. Int. J. Environ. Res. Public Health; 2017; 14, 840. [DOI: https://dx.doi.org/10.3390/ijerph14080840]
3. Deng, J.-Y.; He, Y.; Dai, M. Evaluation of the outdoor thermal environment for three typical urban forms in Nanjing, China. Build. Environ.; 2023; 238, 110358. [DOI: https://dx.doi.org/10.1016/j.buildenv.2023.110358]
4. Ramyar, R.; Ramyar, A.; Kialashaki, Y.; Bryant, M.; Ramyar, H. Exploring reconfiguration scenarios of high-density urban neighborhoods on urban temperature–The case of Tehran (Iran). Urban For. Urban Green.; 2019; 44, 126398. [DOI: https://dx.doi.org/10.1016/j.ufug.2019.126398]
5. Zhou, S.; Wang, Y.; Jia, W.; Wang, M.; Wu, Y.; Qiao, R.; Wu, Z. Automatic responsive-generation of 3D urban morphology coupled with local climate zones using generative adversarial network. Build. Environ.; 2023; 245, 110855. [DOI: https://dx.doi.org/10.1016/j.buildenv.2023.110855]
6. Wu, A.N.; Biljecki, F. InstantCITY: Synthesising morphologically accurate geospatial data for urban form analysis, transfer, and quality control. ISPRS J. Photogramm. Remote Sens.; 2023; 195, pp. 90-104. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2022.11.005]
7. Zhang, W.; Ma, Y.; Zhu, D.; Dong, L.; Liu, Y. Metrogan: Simulating urban morphology with generative adversarial network. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining; Washington, DC, USA, 14–18 August 2022; pp. 2482-2492.
8. Isola, P.; Zhu, J.-Y.; Zhou, T.; Efros, A.A. Image-to-image translation with conditional adversarial networks. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; Honolulu, HI, USA, 21–26 July 2017; pp. 1125-1134.
9. Wang, T.-C.; Liu, M.-Y.; Zhu, J.-Y.; Tao, A.; Kautz, J.; Catanzaro, B. High-resolution image synthesis and semantic manipulation with conditional gans. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; Salt Lake City, UT, USA, 18–23 June 2018; pp. 8798-8807.
10. Huang, S.; He, C.; Cheng, R. SoloGAN: Multi-domain multimodal unpaired image-to-image translation via a single generative adversarial network. IEEE Trans. Artif. Intell.; 2022; 3, pp. 722-737. [DOI: https://dx.doi.org/10.1109/TAI.2022.3187384]
11. Mohamed, O.Y.A.; Zahidi, I. Artificial intelligence for predicting urban heat island effect and optimising land use/land cover for mitigation: Prospects and recent advancements. Urban Clim.; 2024; 55, 101976. [DOI: https://dx.doi.org/10.1016/j.uclim.2024.101976]
12. Chun, P.j.; Izumi, S.; Yamane, T. Automatic detection method of cracks from concrete surface imagery using two-step light gradient boosting machine. Comput.-Aided Civ. Infrastruct. Eng.; 2021; 36, pp. 61-72. [DOI: https://dx.doi.org/10.1111/mice.12564]
13. Yu, Z.; Wang, J.; Yang, X.; Ma, J. Superpixel-Based Style Transfer Method for Single-Temporal Remote Sensing Image Identification in Forest Type Groups. Remote Sens.; 2023; 15, 3875. [DOI: https://dx.doi.org/10.3390/rs15153875]
14. Wang, X.; Yan, H.; Huo, C.; Yu, J.; Pant, C. Enhancing Pix2Pix for remote sensing image classification. Proceedings of the 2018 24th International Conference on Pattern Recognition (ICPR); Beijing, China, 20–24 August 2018; pp. 2332-2336.
15. Lin, Y.; Song, M. Exploring the Potential of Generative Adversarial Networks in Enhancing Urban Renewal Efficiency. Sustainability; 2024; 16, 5768. [DOI: https://dx.doi.org/10.3390/su16135768]
16. Wang, J.; Yang, Y.; Chen, Y.; Han, Y. LighterGAN: An Illumination Enhancement Method for Urban UAV Imagery. Remote Sens.; 2021; 13, 1371. [DOI: https://dx.doi.org/10.3390/rs13071371]
17. Yu, H.; Chen, H.; Zhang, L. An Assessment of the Map-Style Influence on Generalization with CycleGAN: Taking Line Features as an Example. ISPRS Int. J. Geo-Inf.; 2024; 13, 418. [DOI: https://dx.doi.org/10.3390/ijgi13120418]
18. Zhu, M.; Gong, S.; Qian, Z.; Zhang, L. A brief review on cycle generative adversarial networks. Proceedings of the 7th IIAE International Conference on Intelligent Systems and Image Processing (ICISIP); Taipei, Taiwan, 5–9 September 2019; pp. 235-242.
19. Sohn, K.; Berthelot, D.; Carlini, N.; Zhang, Z.; Zhang, H.; Raffel, C.A.; Cubuk, E.D.; Kurakin, A.; Li, C.-L. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Adv. Neural Inf. Process. Syst.; 2020; 33, pp. 596-608.
20. Huang, J.; Huang, X.; Zhang, T.; Jiang, J.; Zhou, H. Asymmetric Two-Stage CycleGAN for Generation of Faces With Shadow Puppet Style. IEEE Access; 2022; 10, pp. 132863-132874. [DOI: https://dx.doi.org/10.1109/ACCESS.2022.3230759]
21. Qin, J.; Zhu, M.; Liu, Y.; Zeng, G.; Xu, B. A cupping spots image enhancement algorithm based on LAA-CycleGAN. Signal Image Video Process.; 2024; 18, pp. 3155-3162. [DOI: https://dx.doi.org/10.1007/s11760-023-02978-1]
22. Tran, N.-T.; Tran, V.-H.; Nguyen, N.-B.; Nguyen, T.-K.; Cheung, N.-M. On data augmentation for GAN training. IEEE Trans. Image Process.; 2021; 30, pp. 1882-1897. [DOI: https://dx.doi.org/10.1109/TIP.2021.3049346] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33428571]
23. Abdelmotaal, H.; Abdou, A.A.; Omar, A.F.; El-Sebaity, D.M.; Abdelazeem, K. Pix2pix conditional generative adversarial networks for scheimpflug camera color-coded corneal tomography image generation. Transl. Vis. Sci. Technol.; 2021; 10, 21. [DOI: https://dx.doi.org/10.1167/tvst.10.7.21]
24. Wang, G.; Shi, H.; Chen, Y.; Wu, B. Unsupervised image-to-image translation via long-short cycle-consistent adversarial networks. Appl. Intell.; 2023; 53, pp. 17243-17259. [DOI: https://dx.doi.org/10.1007/s10489-022-04389-0]
25. Senapati, R.K.; Satvika, R.; Anmandla, A.; Ashesh Reddy, G.; Anil Kumar, C. Image-to-image translation using Pix2Pix GAN and cycle GAN. Data Intelligence and Cognitive Informatics; Springer: Singapore, 2023; pp. 573-586.
26. Stewart, I.D.; Oke, T.R. Local climate zones for urban temperature studies. Bull. Am. Meteorol. Soc.; 2012; 93, pp. 1879-1900. [DOI: https://dx.doi.org/10.1175/BAMS-D-11-00019.1]
27. Demuzere, M.; Bechtel, B.; Middel, A.; Mills, G. Mapping Europe into local climate zones. PLoS ONE; 2019; 14, e0214474. [DOI: https://dx.doi.org/10.1371/journal.pone.0214474] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31017939]
28. Demuzere, M.; Kittner, J.; Bechtel, B. LCZ Generator: A web application to create Local Climate Zone maps. Front. Environ. Sci.; 2021; 9, 637455. [DOI: https://dx.doi.org/10.3389/fenvs.2021.637455]
29. Aslam, A.; Rana, I.A. The use of local climate zones in the urban environment: A systematic review of data sources, methods, and themes. Urban Clim.; 2022; 42, 101120. [DOI: https://dx.doi.org/10.1016/j.uclim.2022.101120]
30. Sundar, G.; Patchaiammal, P.; Venkateshwarulu, B.V.; Kumar, T.P.; Rajamannar, K.; Tripathi, R.K. Integrative Remote Sensing Approaches Using Generative Adversarial Networks for Urban Heat Island Analysis and Mitigation. Remote Sens. Earth Syst. Sci.; 2024; 7, pp. 681-698. [DOI: https://dx.doi.org/10.1007/s41976-024-00156-6]
31. Wu, A.N.; Stouffs, R.; Biljecki, F. Generative Adversarial Networks in the built environment: A comprehensive review of the application of GANs across data types and scales. Build. Environ.; 2022; 223, 109477. [DOI: https://dx.doi.org/10.1016/j.buildenv.2022.109477]
32. Wang, Z.; Peng, Y.; Li, Y.; Xilin, Z.; Xie, Y. Exploration of Influencing Factors of Land Surface Temperature in Cities Within the Beijing–Tianjin–Hebei Region Based on Local Climate Zone Scheme. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.; 2024; 17, pp. 9728-9744. [DOI: https://dx.doi.org/10.1109/JSTARS.2024.3396167]
33. Yang, X.; Xu, S.; Peng, L.L.H.; Chen, Y.; Yao, L. General air temperature and humidity features of local climate zones: A multi-city observational study in eastern China. Urban Clim.; 2023; 51, 101652. [DOI: https://dx.doi.org/10.1016/j.uclim.2023.101652]
34. Zhang, J.; Liu, N.; Wang, S. Generative design and performance optimization of residential buildings based on parametric algorithm. Energy Build.; 2021; 244, 111033. [DOI: https://dx.doi.org/10.1016/j.enbuild.2021.111033]
35. Ao, Y.; Yang, H.; Wei, G.; Ji, B.; Shi, W.; Jiang, Z. An Impoved U-GAT-IT model for Enhancing the Realism of Simulated Ultrasound Images. Proceedings of the 2021 International Conference on Electronic Information Engineering and Computer Science (EIECS); Changchun, China, 23–26 September 2021; pp. 336-340.
36. Creswell, A.; White, T.; Dumoulin, V.; Arulkumaran, K.; Sengupta, B.; Bharath, A.A. Generative adversarial networks: An overview. IEEE Signal Process. Mag.; 2018; 35, pp. 53-65. [DOI: https://dx.doi.org/10.1109/MSP.2017.2765202]
37. Bechtel, B.; Alexander, P.J.; Beck, C.; Böhner, J.; Brousse, O.; Ching, J.; Demuzere, M.; Fonte, C.; Gál, T.; Hidalgo, J. Generating WUDAPT Level 0 data–Current status of production and evaluation. Urban Clim.; 2019; 27, pp. 24-45. [DOI: https://dx.doi.org/10.1016/j.uclim.2018.10.001]
38. Wang, R.; Ren, C.; Xu, Y.; Lau, K.K.-L.; Shi, Y. Mapping the local climate zones of urban areas by GIS-based and WUDAPT methods: A case study of Hong Kong. Urban Clim.; 2018; 24, pp. 567-576. [DOI: https://dx.doi.org/10.1016/j.uclim.2017.10.001]
39. Zhou, X.; Okaze, T.; Ren, C.; Cai, M.; Ishida, Y.; Mochida, A. Mapping local climate zones for a Japanese large city by an extended workflow of WUDAPT Level 0 method. Urban Clim.; 2020; 33, 100660. [DOI: https://dx.doi.org/10.1016/j.uclim.2020.100660]
40. Shao, S.; Wang, P.; Yan, R. Generative adversarial networks for data augmentation in machine fault diagnosis. Comput. Ind.; 2019; 106, pp. 85-93. [DOI: https://dx.doi.org/10.1016/j.compind.2019.01.001]
41. Toda, R.; Teramoto, A.; Kondo, M.; Imaizumi, K.; Saito, K.; Fujita, H. Lung cancer CT image generation from a free-form sketch using style-based pix2pix for data augmentation. Sci. Rep.; 2022; 12, 12867. [DOI: https://dx.doi.org/10.1038/s41598-022-16861-5] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35896575]
42. Sandfort, V.; Yan, K.; Pickhardt, P.J.; Summers, R.M. Data augmentation using generative adversarial networks (CycleGAN) to improve generalizability in CT segmentation tasks. Sci. Rep.; 2019; 9, 16884. [DOI: https://dx.doi.org/10.1038/s41598-019-52737-x]
43. Aggarwal, A.; Mittal, M.; Battineni, G. Generative adversarial network: An overview of theory and applications. Int. J. Inf. Manag. Data Insights; 2021; 1, 100004. [DOI: https://dx.doi.org/10.1016/j.jjimei.2020.100004]
44. Gonog, L.; Zhou, Y. A review: Generative adversarial networks. Proceedings of the 2019 14th IEEE Conference on Industrial Electronics and Applications (ICIEA), Xi’an; China, 19–21 June 2019; pp. 505-510.
45. Chen, Y.-C.; Shibata, H.; Chen, L.-H.; Takama, Y. Synthesis of Comic-Style Portraits Using Combination of CycleGAN and Pix2Pix. J. Adv. Comput. Intell. Intell. Inform.; 2024; 28, pp. 1085-1094. [DOI: https://dx.doi.org/10.20965/jaciii.2024.p1085]
46. Zhu, W.; Wang, Z.; Li, Q.; Zhu, C. A Method of Enhancing Silk Digital Printing Color Prediction through Pix2Pix GAN-Based Approaches. Appl. Sci.; 2024; 14, 11. [DOI: https://dx.doi.org/10.3390/app14010011]
47. Sun, J.; Du, Y.; Li, C.; Wu, T.-H.; Yang, B.; Mok, G.S. Pix2Pix generative adversarial network for low dose myocardial perfusion SPECT denoising. Quant. Imaging Med. Surg.; 2022; 12, 3539. [DOI: https://dx.doi.org/10.21037/qims-21-1042]
48. Lin, E. Comparative Analysis of Pix2Pix and CycleGAN for image-to-image translation. Highlights Sci. Eng. Technol.; 2023; 39, pp. 915-925. [DOI: https://dx.doi.org/10.54097/hset.v39i.6676]
49. Haiderbhai, M.; Ledesma, S.; Lee, S.C.; Seibold, M.; Fürnstahl, P.; Navab, N.; Fallavollita, P. Pix2xray: Converting RGB images into X-rays using generative adversarial networks. Int. J. Comput. Assist. Radiol. Surg.; 2020; 15, pp. 973-980. [DOI: https://dx.doi.org/10.1007/s11548-020-02159-2] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32342258]
50. Li, Z.; Guan, B.; Wei, Y.; Zhou, Y.; Zhang, J.; Xu, J. Mapping new realities: Ground truth image creation with pix2pix image-to-image translation. arXiv; 2024; arXiv: 2404.19265
51. Cira, C.-I.; Manso-Callejo, M.-a.; Alcarria, R.; Fernandez Pareja, T.; Bordel Sanchez, B.; Serradilla, F. Generative learning for postprocessing semantic segmentation predictions: A lightweight conditional generative adversarial network based on Pix2pix to improve the extraction of road surface areas. Land; 2021; 10, 79. [DOI: https://dx.doi.org/10.3390/land10010079]
52. Hölscher, D.; Reich, C.; Gut, F.; Knahl, M.; Clarke, N. Pix2Pix hyperparameter optimisation prediction. Procedia Comput. Sci.; 2023; 225, pp. 1009-1018. [DOI: https://dx.doi.org/10.1016/j.procs.2023.10.088]
53. Henry, J.; Natalie, T.; Madsen, D. Pix2pix gan for image-to-image translation. Res. Gate Publ.; 2021; pp. 1-5. [DOI: https://dx.doi.org/10.13140/RG.2.2.32286.66887]
54. Tirel, L.; Ali, A.M.; Hashim, H.A. Novel hybrid integrated Pix2Pix and WGAN model with Gradient Penalty for binary images denoising. Syst. Soft Comput.; 2024; 6, 200122. [DOI: https://dx.doi.org/10.1016/j.sasc.2024.200122]
55. Zhao, D.; Xu, L.; Chen, L.; Yan, Y.; Duan, L.-Y. Mask-Pix2Pix Network for Overexposure Region Recovery of Solar Image. Adv. Astron.; 2019; 2019, 5343254. [DOI: https://dx.doi.org/10.1155/2019/5343254]
56. Salehi, P.; Chalechale, A. Pix2pix-based stain-to-stain translation: A solution for robust stain normalization in histopathology images analysis. Proceedings of the 2020 International Conference on Machine Vision and Image Processing (MVIP); Qom, Iran, 18–20 February 2020; pp. 1-7.
57. Chu, C.; Zhmoginov, A.; Sandler, M. Cyclegan, a master of steganography. arXiv; 2017; arXiv: 1712.02950
58. Welander, P.; Karlsson, S.; Eklund, A. Generative adversarial networks for image-to-image translation on multi-contrast mr images-a comparison of cyclegan and unit. arXiv; 2018; arXiv: 1806.07777
59. Kaneko, T.; Kameoka, H. Cyclegan-vc: Non-parallel voice conversion using cycle-consistent adversarial networks. Proceedings of the 2018 26th European Signal Processing Conference (EUSIPCO); Rome, Italy, 3–7 September 2018; pp. 2100-2104.
60. Ghosh, B.; Dutta, I.K.; Carlson, A.; Totaro, M.; Bayoumi, M. An empirical analysis of generative adversarial network training times with varying batch sizes. Proceedings of the 2020 11th IEEE Annual Ubiquitous Computing, Electronics & Mobile Communication Conference (UEMCON); New York, NY, USA, 28–31 October 2020; pp. 0643-0648.
61. Li, C.; Wand, M. Precomputed real-time texture synthesis with markovian generative adversarial networks. Computer Vision –ECCV 2016, Proceedings of the Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, 11–14 October, Proceedings, Part III; Springer: Berlin/Heidelberg, Germany, 2016; pp. 702-716.
62. Wang, Y.; Zheng, H.-T.; Chen, W.; Zhang, R. LambdaGAN: Generative adversarial nets for recommendation task with lambda strategy. Proceedings of the 2019 International Joint Conference on Neural Networks (IJCNN); Budapest, Hungary, 14–19 July 2019; pp. 1-8.
63. Salehi, P.; Chalechale, A.; Taghizadeh, M. Generative adversarial networks (GANs): An overview of theoretical model, evaluation metrics, and recent developments. arXiv; 2020; arXiv: 2005.13178
64. Li, W.; Ding, W.; Sadasivam, R.; Cui, X.; Chen, P. His-GAN: A histogram-based GAN model to improve data generation quality. Neural Netw.; 2019; 119, pp. 31-45. [DOI: https://dx.doi.org/10.1016/j.neunet.2019.07.001] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31376636]
65. Lupión, M.; Sanjuan, J.F.; Ortigosa, P.M. Using a Multi-GPU node to accelerate the training of Pix2Pix neural networks. J. Supercomput.; 2022; 78, pp. 12224-12241. [DOI: https://dx.doi.org/10.1007/s11227-022-04354-1]
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Details
; Xiong, Ziheng 1 ; Zhao, Jiayu 1 ; Zhou, Shiqi 2 ; Wang, Yuankai 3 ; Rana Muhammad Adnan Ikram 1
; Wang, Lie 4 ; Tan, Soon Keat 5
1 College of Architecture and Urban Planning, Guangzhou University, Guangzhou 510006, China;
2 College of Design and Innovation, Tongji University, Shanghai 200093, China;
3 Department of Urban Planning and Design, The University of Hong Kong, Pok Fu Lam Road, Hong Kong, China;
4 Art School, Hunan University of Information Technology, Changsha 410151, China
5 School of Civil and Environmental Engineering, Nanyang Technological University, Singapore 639798, Singapore;