Combining generative adversarial networks with urban noise mapping

Junpai Chen; Qiuya Xiang

doi:10.1515/noise-2025-0021

Article Open Access

Combining generative adversarial networks with urban noise mapping

Junpai Chen and Qiuya Xiang

Published/Copyright: November 24, 2025

Published by

Become an author with De Gruyter Brill

Author Information

From the journal Noise Mapping Volume 12 Issue 1

Abstract

Urban traffic noise is associated with the health and living environment quality of residents. As urbanization and population density continually increase, it is vital to understand and predict the impact of urban design behavior on urban traffic noise. Despite the current progress has been made in modeling traffic noise using limited land use types, understanding the complex relationship between various land uses and traffic noise remains challenging for stakeholders. This study used generative adversarial networks with Hong Kong one-hour peak traffic noise map to predict urban traffic noise. The applicability of the training model was evaluated through accuracy analysis and validation. The validated model was used to generate the predicted noise map in multiple scenario experiments by adjusting controlled variables. This approach explores how land use changes effect the noise level, with scenario experiments highlighting both effective strategies and areas requiring further validation.

Keywords: urban noise; noise mapping; generative adversarial networks; machine learning

1 Introduction

1.1 Research background

Noise pollution poses a serious threat to the health of urban residents [1]. Urban populations are increasing and transportation networks are expanding [1,2]. As a result, traffic noise has emerged as the primary source of noise pollution in urban environments [3], as already observed in Europe [4]. Therefore, traffic noise has become a key issue in multiple fields, such as urban planning, public health, and acoustic engineering [5,6]. Although considering noise impact during early planning stages can reduce costs [7], this factor is still overlooked in initial project phases [1]. Early efforts had attempted to monitor the traffic noise levels based on noise mapping or noise factor analysis [8,9], to evaluate or manage noise levels within cities. Relevant work includes the noise model development [10], optimization [11], simulation [12], evaluating [13], and validation of noise models [14].

However, evaluating noise levels in large cities remains a challenging task [15]. Some studies have analyzed noise distribution levels by using noise maps [7,13]. Traditional models represented by UK Calculation of Road Traffic Noise [16], mathematical-statistical models [17], spatial model [13], and machine learning represented by multilayer perceptron or artificial neural network [15,18] have been introduced into different types of traffic noise research to achieve traffic noise assessment and prediction. However, due to the limitations of grid-based noise mapping methods and simulation software in early design [7,19], it was restricted to evaluating traffic noise impact at ex-ante stage. These approaches are limited in capturing traffic noise variations during the process of land-use changes. Furthermore, considering the current demand for efficiency and interactivity of method [1], it is necessary to improve the existing urban noise assessment and prediction approaches within urban design process.

1.2 Noise spatial distribution and land use

Computer science advancements have accelerated the development of spatial-based traffic noise modeling and prediction. Previous studies mapped the relationship between traffic noise and urban form using noise level matrices [20] or geographic information modeling [21]. For spatial correlation modeling, most studies used morphological variables within spatial grids [22]. As one of the main broads of urban form [23], others employed spatial statistical analyses to explore how urban land use elements affect traffic noise [24]. Moreover, the prior knowledge of the aforementioned studies enabled traffic noise prediction based on urban land use characteristics and parameters. It is considered an approach that allows designers and planners to implement low-cost tests to improve solutions [25].

Considering the correlation between morphological composition of urban land use and urban noise [22], it is necessary to implement noise level prediction based on land use change. Some researchers have tried to apply machine learning to noise simulation based on land use pattern change. For example, Yin et al. explored the impact amplitude of defined point, line, and polygon factors on traffic noise based on multiclass machine learning models, including random forest, extreme gradient boosting (XGB), and neural networks (NNs) [26]. In addition, Fallah-Shorshani et al. introduced the regression model, XGB, and three numerical/acoustic traffic noise models on the California dataset, to implement noise estimation, model validation, and comparative studies [27]. Among the applications of adversarial generative networks (GANs), Yang et al. attempted to develop a fast urban traffic noise mapping technique using the GAN model, which was modularly integrated into a design tool within Rhino/Grasshopper [1]. In brief, various types of machine learning have been widely introduced in traffic noise simulation and prediction. It is necessary to continue improving existing methods or providing better tools to enhance application value.

Due to their strong performance in high-resolution image classification and generation tasks [28,29], GANs are widely used in urban planning, particularly for image-to-image translation tasks [30]. The pix2pix algorithm, based on the conditional GAN model, was introduced by Isola et al. [31], later evolving into pix2pixHD [32]. For example, Liu et al. validated that satellite images can accurately predict noise distribution in GAN-based models [33]. These studies demonstrate the potential of combining GANs with satellite imagery for detailed and iterative urban noise mapping. Hence, we use the GAN structure model for further work.

1.3 Problem statement

Machine learning combined with limited land-use elements has been widely used in studies of building facades and road networks [34]. However, these studies often focus on a single or limited land-use type, lacking a more comprehensive consideration of multiple urban land-use within the urban planning [35]. This limitation hinders in-depth exploration of broader land-use factors and their impact on traffic noise. In addition, most studies still utilize spatial numerical models and traditional machine learning approaches, which restrict their applicability to larger datasets and limit the discovery of potential data patterns. Consequently, research on noise environment remains largely constrained to limited scale or condition, with an insufficient approach for larger-scale noise data analysis [36,37]. Briefly, many studies are restricted by the limitations of existing models and algorithms, which limit their applicability in urban planning and design process.

Urban planning and design processes are inherently iterative and rely on dynamic strategies [38]. Current workflows depend on user–computer interactions and simulation feedback for decision-making [39]. Hence, there is a pressing need for a method capable of delivering interactive, fine-grained scale noise prediction feedback, and allowing users to integrate it into the workflow for further decision-making processes. Although Yang et al. highlighted the reliability of using machine learning for fine-grained and iterative noise prediction [1], there remains a lack of empirical evidence and detailed multifactor analyses. To increase the accuracy and evidence of traffic noise predictions, it is essential to adopt finer-scale models and datasets to extend their application to urban-scale training and prediction. Moreover, the lack of accurate data and appropriate methodologies resulted in a scarcity of interactive tools to support intuitive engagement for non-experts, such as designers, planners, and decision-makers, in understanding traffic noise implications [1]. This issue makes it difficult for stakeholders to understand the potential effects of land-use changes on traffic noise. Therefore, developing an interactive, rapid iterative design process that evaluates the impact of land-use changes on traffic noise during the early stages of the project is crucial. Because it could help mitigate potential noise pollution issues at the early stage, reducing the risk of adverse impacts before project implementation [7]. In general, this aims to propose an approach that allows stakeholders to understand the correlation between urban land use layout and noise in the early stage of urban planning.

2 Method

2.1 Research framework

The proposed research framework is shown in Figure 1. The detailed steps of this study are as follows: The one-hour peak traffic noise (OHPTN) data of Hong Kong was first obtained from the Environmental Protection Department (EPD) of the Hong Kong Special Administrative Region Government, and exported to a OHPTN map with a 4-m resolution during 2020 [40]. Simultaneously, both LiDAR and satellite datasets were analyzed to stack a land use map, such as the building footprint, road networks, water bodies, and normalized difference vegetation index (NDVI) layers. After processing the NDVI layer into a labeled layer, all morphological feature layers were stacked as the land use map at the fine-grain level with a 5-m resolution. Second, the OHPTN map and the land use map were processed and segmented into a training image dataset and a test image dataset. Third, the GAN models were trained using the training image dataset from the OHPTN map and the land use map. The accuracy analysis and validation were performed to select the acceptable model. Finally, multiscenario experiments were implemented based on land use element variable control, to obtain the sensitivity of different elements to noise changes and explore land use solutions to reduce noise.

Figure 1

Research framework.

2.2 Data collection and pre-processing

2.2.1 Traffic noise map

The raw noise map dataset is the open dataset from the Common Spatial Data Infrastructure (CSDI) Portal and EPD [40,41]. This dataset provides an annual average noise level of L 10 (1-h, 4-m height) for Hong Kong in 2020 [42]. L 10 describes the noise level exceeded for 10% of the 1 h period, usually the traffic noise during the hour of peak traffic flow. The OHPTN data for 2020 was downloaded from the CSDI platform in GeoTIFF format from the EPD of the Hong Kong SAR Government and imported into ArcGIS Pro. Using the primary symbology system with a min–max renderer, noise values ranging from 0 to 96.83 dB were linearly rescaled to grayscale values between 0 and 255. Then a OHPTN map was exported at 4 m resolution by using the Hong Kong 1980 coordinate system (EPSG: 2326). The noise map processing was presented via linear workflow in Figure 2.

Figure 2

OHPTN map processing workflow.

2.2.2 Land use map

To process the input map based on the land use of Hong Kong, we combined various data sources, such as LiDAR raster and other open datasets from the CSDI Portal [41]. The georeferencing followed the UTM projection and the Hong Kong 1980 datum to align with TM/ETM+/OLI imagery. Following our previous work process [43], this 11-level dataset were assigned specific colors using an 8-bit RGB system (Figure 3 and Table 1). The results were resampled to generate an 11-level LULC raster grid with a 5-m spatial resolution. Then, according to the color labels in Figure 3, after performing color processing mapping on each data source in the ESRI ArcGIS software, these layers were stacked together and exported as a complete JPG format image.

Figure 3

Land use map processing workflow.

Table 1

Comparison of input and output datasets

	Input dataset	Output dataset
Description	Land use map	OHPTN map
Image data format	33,540 pixels by 23,540 pixels, 1,920 dpi	33,540 pixels by 23,540 pixels, 1,920 dpi
Composition	Building footprint datasets, NDVI datasets, LiDAR datasets	One-hour peak traffic noise maps
Data resource	CSDI, Google Map, Open Street Map, USGS	CSDI
Coordinate	Hong Kong 1980 (EPSG: 2326)	Hong Kong 1980 (EPSG: 2326)
Number of datasets	3,192 (test dataset: 319, Train dataset: 2,873)	3,192 (test dataset: 319, train dataset: 2,873)
Data label number	11	Continuous
Data label content	Building, road infrastructure, water, impermeable surface and bare soil surface, tree (canopy), shrub, grassland, noise enclosure, noise barrier	One-hour peak traffic noise value per pixel

2.2.3 Image processing and segment

Then, two maps were converted into the corresponding paired image dataset for GAN model training. First, we aligned the coordinate systems of the OHPTN output map and the land use input map to the Hong Kong 1980 projection (EPSG: 2326). Both maps were then exported as high-resolution images (33,540 by 23,540 pixels, 1920 dpi). Next, the Python Imaging Library (PIL) [44] was used to partition the large images into smaller paired images. Each segmented image covered an area of approximately 1,000 by 1,000 m (512 by 512 pixels). Each pixel represents an actual distance of 1.95 m by 1.95 m, where cover area occupied by each pixel is approximately 3.81 m 2 . To reduce training costs, any paired images located outside of administrative boundaries were excluded. Then 90% of the paired images were used for model training (2,873 paired images), and 10% of the paired images were used as the test image datasets (319 paired images) for analyzing the accuracy of the model (Figure 4).

Figure 4

Distribution map of the train image dataset and test image dataset.

2.3 GAN model structure and training

The structure of GAN includes two main components: a generator ( G ) that creates synthetic samples and a discriminator ( D ) that distinguishes between real and generated data (Figure 5). During training, the generator improves by minimizing the difference between its generated data and real data, while the discriminator adjusts to better differentiate between the two. The training process continues until the discriminator cannot reliably distinguish between real and fake data [45]. In this study, the GAN-based model uses land use maps as inputs and OHPTN images as outputs for training. The pix2pixHD model was chosen, with hyperparameters set according to previous studies [46,47]. After 1,000 epochs of stable training, the output from the model reached stable patterns. The training duration of 1,000 epochs was chosen based on previous works [1,46], as the focus was not on optimizing the network’s performance but rather on achieving a reliable result for this specific application.

Figure 5

Structure of GAN model.

2.4 Accuracy analysis and validation

A series of quantitative methods was applied to measure the performance of the model throughout various stages of training. To overcome overfitting, this study incorporated the enhanced loss function (LOSS) derived from the pix2pixHD framework [48]. The following equations detail how LOSS was computed for both the generator and discriminator components during training:

ℒ FM ( G , D k ) = E ( s , x ) ∑ i = 1 T 1 N i ∥ D k ( i ) ( s , x ) − D k ( i ) ( s , G ( s ) ) ∥ 1

min G ∑ k = 1,2,3 max D 1 , D 2 , D 3 ℒ GAN ( G , D k ) + λ ∑ k = 1,2,3 ℒ FM ( G , D k ) ,

where T represents the number of layers involved, and N i indicates the count of individual components in the i th layer. The parameter λ regulates the influence between the adversarial and feature-matching loss components.

To assess model performance quantitatively over the course of training, four evaluation indices were utilized. As suggested in prior studies, these consist of inception score (IS) [49], Frechet inception distance (FID) [50], learned perceptual image patch similarity (LPIPS) [51], and structural similarity index metric (SSIM) [52]. The accuracy assessment was carried out over a complete training of 1,000 epochs, with evaluations performed at each epoch. The calculation of aforementioned metrics are as follows:

IS = exp [ E z ∼ p ( z ) [ D ( p ( y ∣ g ( z ) ) ‖ p ( y ) ) ] ] FID = ∣ μ x − μ y ∣ 2 − T r ( ∑ x + ∑ y − 2 ∑ x ∑ y ) LPIPS ( X , Y ) = 1 N ∑ i = 1 N ∣ ϕ i ( x ) − ϕ i ( y ) ∣ 2 SSIM ( X , Y ) = 2 μ X μ Y + ( K 1 L ) 2 μ X 2 + μ Y 2 + K 1 L α ⋅ 2 σ X σ Y + ( K 2 L ) 2 σ X 2 + σ Y 2 + K 2 L β ⋅ 2 σ X Y + K 3 2 σ X + σ Y + K 3 γ .

The metric calculations were implemented using the TorchMetric library [53]. Specifically, modules corresponding to Frechet inception distance, inception score, LPIPS, and SSIM were employed. Furthermore, to further quantify predictive uncertainty and the confidence level of reported noise reductions, we compared the generated results with the reference OHPTN data on the test datasets using a consistent grayscale-to-decibel mapping and paired statistics within identical pixel supports. Grayscale values were converted to sound levels (dB) based on Figure 2. For each valid pixel ( i = 1 , … , N ), white pixels were excluded), residuals were defined as follows:

e i = y ˆ i − y i ,

where y ˆ i is the predicted noise level (in dB) and y i is the reference OHPTN value. The mean error (ME), mean absolute error, and residual standard deviation ( σ ) as follows:

ME = 1 N ∑ i = 1 N e i MAE = 1 N ∑ i = 1 N ∣ e i ∣ σ = 1 N − 1 ∑ i = 1 N ( e i − ME ) 2 .

For a region of interest (ROI), the standard error (SE) of the mean change ( Δ ) is approximated as follows:

SE = σ N .

Here, the two-sided 95% confidence interval for mean change is given by:

CI 95 % = Δ ± 1.96 × SE .

Equivalently, the minimum ROI size required to reliably detect a change of magnitude ( ∣ Δ ∣ ) at the 95% level is as follows:

N ∗ ( ∣ Δ ∣ ) = 1.96 , σ ∣ Δ ∣ 2 .

To avoid distributional assumptions and to accommodate spatial dependence, 95% confidence intervals for image-level summary errors (MAE) were additionally estimated.

2.5 Scenario experiment

The scenario-based simulation was systematically controlled relying on the one-factor-at-a-time approach [54]. We adjusted individual and combined variables to investigate the sensitivity between peak-hour traffic noise and land use factors. As in our previous work [43], we used Adobe Photoshop to manually change color values and pixel quantities in a predefined image. This process enabled the creation of a modified map representing urban environmental elements, influenced by the controlled scenario variables. The same software and process were applied consistently across all experimental conditions. Finally, the prediction results were recalculated based on the chosen model and the adjusted land-use map. Unless otherwise specified, all the sample areas were randomly selected from the test dataset, and subsequent scene simulation studies were conducted under the condition of ensuring there was sufficient available space. The global grayscale values from the prediction results and the original map were used to analyze the trend of traffic noise variations by calculating the changes in grayscale value. Based on the legend of grayscale-noise level values in Figure 2, each grayscale unit corresponds to 0.38 dB . Thus, the noise level reduction between the two layouts can be expressed as follows:

N R = Δ f ( x ) × 0.38 ,

where N R is the noise level reduction in decibel (dB) and Δ f ( x ) is the difference in grayscale values between scenarios. Higher grayscale value reduction represents a higher reduction of the noise level value.

3 Results

3.1 Accuracy analysis of training results

3.1.1 Quantitative analysis results

The model was trained for 1,000 rounds on a consumer-grade device with NVIDIA GeForce RTX 3060Ti, Intel(R) Core(TM) i7-12700F, and 16 GB RAM, with a total usage of 53 h and 19 min. Figure 6 shows the values of the four metrics for the model across the training and test sets over 1,000 epochs. Based on the accuracy results, Epoch 990 was selected as the preferred model for subsequent validation and scenario simulation experiments. Epoch 990’s metric performance makes it an acceptable deployment option.

Figure 6

Results of loss function and four metrics. (a) Loss function results, (b) test set accuracy results, and (c) train set accuracy results.

3.1.2 GAN model performance validation

From the results in the previous section, epoch 990 was identified as the model for the scenario simulation experiment in the following section. To ensure the quantitative accuracy of the selected model, additional retraining was implemented. The generated image performance from the selected model was compared across different image categories (Figure 7). Leveraging methods from prior studies [31,32], the outputs of the model were evaluated against those generated by inverse processes, random noise, and real images. The assessment employed the same four metrics, including IS, FID, LPIPS, and SSIM.

Figure 7

Model performance validation and uncertainty quantification results. (a) Image size 512 pixels * 512 pixels, (b) train/test set accuracy analysis, and (c) uncertainty quantification results.

The model-generated outputs outperformed the inverse and random generation. For the test dataset, the generated results from epoch 990 reduced the FID metric by 98.76% in both the inverted image set and noise image set. It increased the SSIM metric by 262.79% in the inverted image set and 5151.28% in the noise image set. The LPIPS metric also improved by 92.02% in the inverted image set and 95.52% in the noise image set. Given the results of the training dataset, the generated results from epoch 990 reduced the FID metric by 99.59% in the inverted image set and 99.60% in the noise image set. It further increased the SSIM metric by 303.31% in the inverted image set and 5233.51% in the noise image set, and it also increased the LPIPS metric by 95.43% in the inverted image set and 97.43% in the noise image set. In summary, epoch 990 was selected as the model for application in subsequent scenario experimental processes. For the results of uncertainty quantification (Figure 7(c)), the model yielded a mean MAE of 2.99 dB ( CI 95 % is 2.88–3.09 dB). Paired bootstrap analysis further indicates that reductions as small as 1 dB can be considered statistically significant at the 95% confidence level in typical regions of interest.

3.2 GAN-assisted urban planning factor scenario

3.2.1 Green space scenario

Previous studies have confirmed that green space is related to urban traffic noise levels [55–57]. Hence, it is crucial to explore the influence of traffic noise on green space distribution under human intervention. A sample area was selected for four experimental scenarios (see Figure 8, Line 1). This experiment hypothesizes that vacant impervious surfaces or bare soil will be converted into more green spaces. To evaluate the effects of design interventions, stakeholders need to understand the layout, coverage, and composition of green spaces. Therefore, the simulated scenarios for green space expansion have been classified into four scenarios. The results based on epoch 990 are shown in Figures 8 and 9.

Figure 8

Green space scenario results part 1.

Figure 9

Green space scenario results part 2.

Scenario 1: The objective in this scenario is to identify the most effective layout to achieve the lowest noise levels for various single-vegetation types under a constant coverage growth rate. In this case, the greening coverage growth rate is set at 1% of the sample (2,622 pixels, 9,989.82 m 2 ). The mixed grassland layout led to the greatest reduction in noise levels ( − 3.49 % , − 1.23 dB). For the shrub type, a distributed block layout produced the highest noise reduction ( − 3.29 % , − 1.16 dB). For tree or forest canopy areas, distributed strip layout was most effective, lowering noise levels by − 4.25 % ( − 1.50 dB).

Scenario 2: Building on the layout results from Scenario 1, this scenario aims to seek growth coverage rates for various single-vegetation types to minimize noise levels. The proposed green space layouts and vegetation combinations from Scenario 1 included three categories: 100% grassland, 100% shrubs, and 100% tree canopy. These categories were examined under five different coverage growth rates: 0.2% (525 pixels, 2000.25 m 2 ), 0.4% (1,049 pixels, 3,996.69 m 2 ), 0.6% (1,573 pixels, 5,993.13 m 2 ), 0.8% (2,098 pixels, 7,993.38 m 2 ), and 1% (2,622 pixels, 9,989.82 m 2 ), as shown in Figure 6. Under specific layout conditions, a growth rate of 0.8% of the tree type, leading to maximized reduction of noise levels ( − 4.66 % , − 1.65 dB). For the grassland type, a growth rate of 0.4% produced the highest reduction in noise levels ( − 3.7 % , − 1.31 dB). At a growth rate of 0.2%, the shrub type provides the greatest decrease in noise level ( − 3.53 % , − 1.25 dB).

Scenarios 3 and 4: These two scenarios aim to identify the most effective combinations of green vegetation that lead to the largest noise reduction. In these cases, the green space layouts and coverage growth rates previously identified were applied to the sample areas. Figure 9 presents the outcomes for different pairwise vegetation combinations under optimal configurations. In the scenario with the best grassland layout, the best double pairing was a grassland and tree(canopy) mix with a 2:8 ratio, which resulted in a noise level reduction of approximately − 3.80 % ( − 1.34 dB). For the shrub layout, the optimal double pairing was shrub and grassland at a 4:6 ratio, leading to a noise reduction of roughly − 3.59 % ( − 1.27 dB). For tree and canopy areas, the best double pairing was tree and grassland in a 6:4 ratio, which reduced the noise by about − 4.93 % ( − 1.75 dB). Further, the best triple pairing was tree dominant with grass, shrub, and tree in a 1:1:8 ratio, which reduced the noise by about − 4.97 % ( − 1.76 dB).

3.2.2 Building scenario

Previous studies have partially examined the influence of buildings on traffic noise in densely populated areas [20,58]. These scenarios assumed building construction on vacant impervious land or bare soil. Stakeholders were tasked with creating design proposals for these areas. The layout and coverage of the building footprints were taken into account, and several experiments were conducted under varying controlled conditions (Figure 10).

Figure 10

Building scenario results.

Scenario 5: In this case, two sites were chosen to represent low-density and high-density zones. Each site covered approximately 1% of the total area (2,622 pixels, 9989.82 m 2 ) for experimentation. With a controlled building coverage rate of 40% (1,049 pixels, 3996.69 m 2 ), the goal of this experiment was to identify the most effective layout in reducing noise levels across different layout types. The building designs included chessboard, block, strip, compact, and mixed layouts, implemented in both areas. The findings revealed that the mixed layout was the most effective in the high-density area, achieving a noise reduction of approximately − 4.68 % ( − 1.65 dB). For the low-density area, the chessboard layout proved most effective, reducing the noise level by about − 11.54 % ( − 3.57 dB).

Scenario 6: With the results from scenario 5, this scenario aimed to assess the impact of various building coverage ratios (BCR) on noise reduction. The BCR levels included 40, 50, 60, 70, and 80%. Based on the predictions from epoch 990 shown in Figure 10, an 80% BCR led to the greatest reduction in noise levels ( − 4.85 % , − 1.72 dB) in the high-density area, while a 70% BCR produced the most significant noise reduction ( − 11.95 % , − 3.69 dB) in the low-density area.

3.2.3 Road scenario

Road networks directly influence traffic noise [59]. To isolate the pure land use effects, all scenario experiments were conducted under the assumption that annual traffic volumes and vehicle mix remain constant, thus excluding induced human activity due to road network modifications. Two sites in Hong Kong were selected: Site A, located in the urban core with a high-density grid pattern, and Site B, representing a radial arterial layout on the countryside. The goal of this section is to assess how variations in road grid structure and width impact noise distributions when traffic conditions themselves are held fixed.

Scenario 7: The objective of this scenario is to determine the pixel width of roads that leads to the highest reduction in noise levels, based on the fixed grid density. Here, “increase” refers to adding 1 pixel (30 m width) of road land on one side of the original road area, while “decrease” refers to reducing 1 pixel (30 m width) of road land on one side of the original road area. The road width was readjusted, and the updated grid was reintroduced into the model to predict changes in noise distribution. According to the model results presented in Figure 11, the optimal road increase solution in site A yielded approximately an − 8.63 % ( − 2.48 dB) decrease in noise level. At point B, the optimal width scheme resulted in a decrease of approximately − 3.79 % ( − 1.46 dB) in noise level (Figure 11).

Figure 11

Road scenario results.

Scenario 8: With the results of scenario 7, this part focuses on modifying the network grid density in two sites to observe the change of noise levels. It aims to connect or remove existing roads on the existing layout to change the structure of the existing road network. Specifically, it involves adding multiple road plots within one or multiple area. The impact of different grid density configurations on noise levels in both sites were presented in Figure 11. The hybrid approach led to a − 6.20 % ( − 1.78 dB) decrease in the highest noise level in site A. On the other hand, the density reduction strategy for one area yielded the lowest increase in noise levels in site B, with a +3.23% (+1.24 dB) increase.

3.2.4 Water scenario

Previous works have examined how water bodies impact the distribution of urban traffic noise [60]. In this section, two sites (Site C and Site D) in Hong Kong were randomly selected (Figure 8, line 1). The experimental hypothesis tests the addition of new artificial water bodies in areas with impermeable surfaces, bare soil, or other vacant land. These types of water bodies include ponds, reservoirs, and channels. The spatial configuration of water was modified to predict its localized effects on traffic noise levels.

Scenarios 9 and 10: Scenario 9 investigated the effects of different layouts on noise levels, while keeping the total area covered by water bodies constant at 1% (2,622 pixels, 9,989.82 m 2 ). The layout configurations included concentrated linear, decentralized linear, concentrated block, decentralized block, mixed, and spot (Figure 8). With the results of Scenario 9, scenario 10 calculated five proportions based on its results, including 0.2% (525 pixels, 2000.25 m 2 ), 0.4% (1,049 pixels, 3996.69 m 2 ), 0.6% (1,573 pixels, 5993.13 m 2 ), 0.8% (2,098 pixels, 7993.38 m 2 ), and 1% (2,622 pixels, 9989.82 m 2 ). Based on the results of scenario 9, the optimal layout for Site D is the mixed form, which increased the lowest noise level by +0.99% (+0.23 dB). The optimal layout for Site C is spot form, which reduced the noise level of − 2.36 % ( − 0.89 dB). In addition, according to the results of scenario 10, the optimal ratio of Site D is 1% with the lowest increased noise level of +0.99% (+0.23 dB). The optimal ratio of Site C is 0.2%, which reduced the noise level by − 2.19 % ( − 0.82 dB).

Scenarios 11 and 12 investigate the layout and proportion of water bodies under local conditions, similar to scenarios 9 and 10. However, the key difference is that both scenarios focus on experimental areas covering 1% of the total sample area (9,989.82 m 2 ). In scenario 11, the modified area was set at 1% (9989.82 m 2 ), with the water body occupying 25% (656 pixels, 2499.36 m 2 ) of the experimental area. Based on the findings from scenario 11, scenario 12 examined five different ratios of coverage, including 5% (131 pixels, 499.11 m 2 ), 10% (262 pixels, 998.22 m 2 ), 15% (394 pixels, 1497.33 m 2 ), 20% (525 pixels, 1996.44 m 2 ), and 25% (656 pixels, 2499.36 m 2 ). The results from scenario 11 showed that the optimal layout for both sites is the concentrated block configuration. It increased the noise levels by +0.69% (+0.16dB) for Site D and reduced the noise levels by − 2.41 % ( − 0.91 dB) for Site C. Scenario 12 determined that the optimal local ratio for both Site C and Site D was 5%, leading to a lowest increase of +0.72% (+0.17dB) for Site D and highest reduction of − 2.51 % ( − 0.94 dB) for Site C (Figure 12).

Figure 12

Water scenario results.

3.2.5 Noise barrier scenario

Noise barriers are commonly employed to mitigate traffic noise in localized areas [61]. Based on the noise barrier data from [62], this scenario implemented 11 different layouts of noise barriers in two site areas. Site E and Site F were selected to represent the center and edge areas. All noise barriers are set on one side or both sides of the primary road with the same width (4-pixel width, 7.8 m width). In addition, Noise barriers that were not completed in 2020 were excluded from the dataset (see Noise Barrier Dataset Statement section of appendix for details). In scenario 13, all layouts were implemented for five different lengths, including 262 pixels (0.1%, 998.22 m 2 ), 525 pixels (0.2%, 2,000.25 m 2 ), 787 pixels (0.3%, 2,998.47 m 2 ), 1,049 pixels (0.4%, 3,996.69 m 2 ), and 1,311 pixels (0.5%, 4,994.91 m 2 ) (Figure 13). The results of scenario 13 showed that the optimal feature of the noise barrier in site E was the dissymmetry layout with 0.4%, which reduced the noise level of site E by − 15.70 % ( − 3.87 dB). The optimal feature of site F was the dissymmetry layout with +3.21%(+1.37 dB).

Figure 13

Noise barrier scenario results.

3.2.6 Noise enclosure scenario

Similar to the last scenario, noise enclosures are also widely used in Hong Kong for traffic noise reduction [63] (Figure 14). Based on the noise barrier data from [64], this scenario implemented four different layouts of noise barriers in two selected sites. In this section, Site E and Site F were still selected to represent the center and edge areas. All noise enclosures are set on the primary road with the same land occupation ratio (0.5%, approximately 1,311 pixels, 4994.91 m 2 ). Based on the results of scenario 14, the optimal feature of noise enclosure in Site E was the discontinuous single-side layout, which reduced the noise level of Site E by − 14.55 % ( − 3.51 dB). The optimal feature of noise enclosure in Site F was the continuous single-sided layout, which only increased the noise level of Site F by +3.74% (+1.59 dB).

Figure 14

Noise enclosure scenario results.

3.2.7 Combined scenario

To effectively assess the interrelationship of multiple land use factors based on GAN models, combined scenarios were implemented under controlled variable conditions. For each scenario, a quantified result with the highest reduction value is selected as a strategy. This scenario assumes the development of a new project on bare soil or other available land. For example, combining the green space scenario and the building scenario means developing green Spaces with tree dominant with 1:1:8 (grass: shrub: tree) ratio and chessboard layout buildings with 70% BCR on the available vacant land simultaneously. As illustrated in Figures 15, 16, 17, all reference combination results were introduced. All experiments were implemented in a single sample area randomly selected from the test set. In general, the implementation of combination strategies led to a reduction in noise levels. The largest decreased value was observed in the combination experiments of the plant, road, noise barrier, and noise enclosure scenarios, with a decrease of − 19.33 % ( − 4.66 dB). Meanwhile, the lowest decrease was observed in the combination experiments of the building, road, and noise enclosure scenarios, with a decrease of − 13.89 % ( − 3.35 dB).

Figure 15

Summary of combination scenario results part 1.

Figure 16

Summary of combination scenario results part 2.

Figure 17

Summary of combination scenario results part 3.

4 Limitations and future work

However, this study still faces several limitations related to data sources, model architecture, and validation methods. Regarding data sources, the current model does not incorporate additional potential driving variables such as climate data, human activity patterns, and anomalous noise events (e.g., construction activities or traffic accidents). These omissions stem from both data accessibility constraints and technological limitations in integrating diverse datasets. Many research teams and local authorities have implemented low-noise road pavements [65–67]. However, such data were excluded due to unclear implementation records, incomplete maintenance information, and inconsistent standards across jurisdictions. Considering that low-noise surfaces can reduce noise levels by approximately 5–8 dB [67], we suggest limiting the prediction percentage error value of this study to between − 0.5 % and +0.5%. Future research should incorporate comprehensive evaluations of infrastructure maintenance data to accurately assess the cost-effectiveness of noise reduction measures.

The model architecture presents additional limitations in generalization. Although this study aimed to enhance forecasting accuracy by refining existing satellite imaging approaches, it did not involve hyperparameter optimization or cross-validation to fine-tune parameters. The research focused primarily on integrating multiple land use layers with the Pix2pixHD model rather than comparing multiple model architectures. This single-model approach may overlook potentially superior deployment options. Furthermore, the spatial and temporal resolution of both input and output data potentially constrains the model’s ability to capture fine-grained noise variations. The validation process, while employing multiple metrics (IS, FID, LPIPS, and SSIM), lacks field validation against actual noise measurements under different urban conditions. These limitations suggest caution when interpreting results and applying the model to other urban contexts without further validation.

In addition, while this study focuses on land use morphology as a key determinant of traffic noise levels, we acknowledge that traffic noise is inherently a multifactorial phenomenon. Other influential variables also play substantial roles, such as traffic flow, vehicle composition, road surface conditions, and meteorological factors [56]. This study emphasizes that land use stems from its long-term planning relevance and the lack of high-resolution, city-scale traffic and environmental datasets in this period. This modeling simplification, although allowing for scenario-based simulation of morphological interventions, may lead to discrepancies with studies that integrate more dynamic or physical noise modeling elements. Future studies should consider integrating spatiotemporally continuous data streams to support more reliable modeling.

Future work should address these limitations through several approaches. First, incorporating additional data sources, including climate variables, human activity patterns, road surface materials, and traffic volume, would enhance the reliability of the model. Second, exploring multimodel comparisons and ensemble methods could identify more accurate and robust prediction results. The development of hybrid models that combine physical noise propagation principles with data-driven approaches might better capture the complex relationships between urban morphology and noise distribution. Third, extending the validation process to include field measurements across diverse urban settings would strengthen reliability in the predictions and identify context-specific adjustments needed for different urban environments.

5 Discussion and conclusion

The proposed approach, implemented with Hong Kong’s traffic noise dataset, has demonstrated its adaptability and effectiveness in predicting fine-grained traffic noise distribution through various scenario experiments. The modification process is not only visually friendly for users but also allows users to implement interaction by directly modifying the land use map using the software. The approach provides a potential solution for effectively predicting traffic noise trends. This enables identification of potential noise issues during the early proposal phase, facilitating proactive mitigation strategies.

While the scenario experiments provided quantitative insights, it is equally important to critically interpret the results in light of known physical and acoustic principles. For example, green space scenarios showed that different green space layouts can significantly influence noise attenuation, such as vertical tree scattering or shrub clustering. These results fit with the conclusions of previous works that green spaces have the structural ability to block or diffuse sound waves [68]. Similarly, the effectiveness of chessboard and mixed building layouts may be attributed to their ability to disrupt continuous sound propagation paths, as mentioned in previous works [68,69]. However, it may contrast with statistical studies reporting that more buildings correlate with higher noise levels [70]. This discrepancy arises because model simulations isolate the spatial and physical blocking effects of building layouts without accounting for induced traffic flows or increased human activity that typically accompany urban densification. In this microscale context, strategic placement of buildings can disrupt sound propagation by introducing reflective and absorptive surfaces, thereby diffusing traffic noise. Nonetheless, these effects could not generalize to broader urban dynamics where densification impacts both noise sources and propagation pathways. The results of the road scenario showed that changes in road width or grid shape have an impact on the noise levels in both high-density and low-density areas. This confirms that well-designed roads have the potential to control traffic noise distribution [71]. Future work can combine the user equilibrium traffic assignment model [72] or other quantitative methods [73] to understand the fine-scale impact of road design changes on the soundscape. Water body scenarios presented more complex patterns. In some cases, noise increased due to surface reflectivity, suggesting that poorly integrated water layouts may amplify rather than absorb noise. The water body results have confirmed the potential feasibility of properly designed water bodies in improving the soundscape [74]. Furthermore, adding noise barriers and enclosures to site F leads to a slight increase in the local noise level (approximately between 1 and 2 dB). This might be caused by reflection effects [75] or other environmental factors [76]. It is suggested that future work should combine field investigation results and models to provide more empirical results. Overall, our results initially confirm the feasibility of using a GAN-based approach to explore the interrelationships between traffic noise and urban land use factors, providing a foundation for evidence-based noise mitigation strategies.

In summary, the scenario experiments partially validate the predictive capabilities of our approach in exploring the relationship between urban land use factors and traffic noise levels, while also highlighting areas where the model results deviate from conventional expectations, thus warranting further investigation and field validation. The model demonstrates the ability to provide quantifiable noise outcomes under different intervention scenarios. Part of our findings align with the previous work. For example, green space design can impact the soundscape [77]. And the predictions of noise barriers (Scenario 13, Map E) confirm that they can mitigate noise levels in high-density areas, as mentioned by the previous work [61]. However, other results challenge conventional understanding, such as the observed increase in local noise levels following noise barrier installation in already quiet areas, or the potential increase in noise levels associated with certain water body configurations. These microscale findings may complement or sometimes contradict broader principles, highlighting the complexity of noise dynamics in urban environments and the value of fine-grained analysis. Nevertheless, the lack of field validation against actual noise measurements suggests caution when interpreting the current results or applying them. The authors believe that this model has the potential to serve as one of the design tools in the early stages of land planning and architectural design, providing a pre-stage prediction result of noise level for planners, architects, and policymakers. However, the main limitation of this method is that it is currently not applicable to the training tasks with sparse datasets or spatially discontinuous datasets. And it is not suitable for high-precision tasks based on strict physical modeling. Any implementation of the aforementioned strategies should consider the potential risks of artificial intelligence and carefully compare the results of this article with other studies.

With continuing urbanization and growing urban populations, addressing traffic noise issues becomes increasingly critical for public health and quality of life. Providing prediction tools during the early stages of urban projects is essential for mitigating and monitoring urban traffic noise trends. Our framework utilizing GANs for traffic noise prediction represents a step toward integrating noise considerations into standard urban planning workflows. By enabling rapid assessment of design alternatives, the approach supports evidence-based decision-making that balances noise reduction with other urban planning objectives. While acknowledging the limitations of the current approach, the demonstrated applicability and the potential of the proposed framework suggest that it could significantly inform stakeholders about the complex relationships between human interventions in land use and urban noise environments. Future refinements to the model and methodology will further enhance its utility as a decision support tool for creating healthier urban soundscapes.

Acknowledgments

We sincerely thank the editors and reviewers for their valuable comments.

Funding information: The authors state no funding involved.
Author contributions: All authors have accepted responsibility for the entire content of this manuscript and consented to its submission to the journal, reviewed all the results and approved the final version of the manuscript. Junpai Chen contributed in conceptualization, data curation, formal analysis, investigation, methodology, software, supervision, validation, visualization, writing of the original draft, and writing (reviewing and editing). Qiuya Xiang contributed in data curation, formal analysis, investigation, methodology, and writing (reviewing and editing).
Conflict of interest: The authors state no conflict of interest.

References

[1] Yang X, Han Z, Lu X, Zhang Y. A rapid approach to urban traffic noise mapping with a generative adversarial network. Appl Acoustics. 2025;228:110268. 10.1016/j.apacoust.2024.110268Search in Google Scholar

[2] Li X, Yuan Y, Liu G, Han Z, Stouffs R. A predictive model for daylight performance based on multimodal generative adversarial networks at the early design stage. Energy Buildings. 2024;305:113876. 10.1016/j.enbuild.2023.113876Search in Google Scholar

[3] U.S. Department of Transportation. 1995. Highway traffic noise analysis and abatement policy and guidance. https://www.nrc.gov/docs/ML1302/ML13023A373.pdf. Accessed on 2024-10-01. Search in Google Scholar

[4] Khomenko S, Cirach M, Barrera-Gömez J, et al. Impact of road traffic noise on annoyance and preventable mortality in European cities: A health impact assessment. Environ Iint. 2022;162:107160. 10.1016/j.envint.2022.107160Search in Google Scholar PubMed

[5] Morillas JMB, Gozalo GR, González DM, Moraga PA, Vílchez-Gómez R. Noise pollution and urban planning. Current Pollution Reports. 2018;4:208–19. 10.1007/s40726-018-0095-7Search in Google Scholar

[6] Alvares-Sanches T, Osborne PE, White PR. Mobile surveys and machine learning can improve urban noise mapping: Beyond a-weighted measurements of exposure. Sci Total Environ. 2021;775:145600. 10.1016/j.scitotenv.2021.145600Search in Google Scholar PubMed

[7] De Vos P, Licitra G. Noise maps in the european union: An overview. In: Gaetano L, editor. Noise map EU models procedures. 1st edn. London: CRC Press; 2013. p. 26. Search in Google Scholar

[8] Yang W, Heee J, Heee C, Cai M. Evaluation of urban traffic noise pollution based on noise maps. Transport Res Part D Transport Environ. 2020;87:102516. 10.1016/j.trd.2020.102516Search in Google Scholar

[9] Cai M, Zou J, Xie J, Ma X. Road traffic noise mapping in guangzhou using gis and gps. Appl Acoustics. 2015;87:94–102. 10.1016/j.apacoust.2014.06.005Search in Google Scholar

[10] Alam P, Ahmad K, Afsar SS, Akhtar N. Noise monitoring, mapping, and modelling studies-a review. J Ecol Eng. 2020;21:82–93. 10.12911/22998993/119804Search in Google Scholar

[11] Cook MR, Gee KL, Transtrum M, Lympany SV, et al. Toward a dynamic national transportation noise map: Modeling temporal variability of traffic volume. J Acoust Soc Am. 2023;154:2950–8. 10.1121/10.0022356Search in Google Scholar PubMed

[12] Vogiatzis K. Airport environmental noise mapping and land use management as an environmental protection action policy tool. the case of the larnaka international airport (cyprus). Sci Total Environ. 2012;424:162–73. 10.1016/j.scitotenv.2012.02.036Search in Google Scholar PubMed

[13] Zhou Z, Zhang M, Gao X, Gao J, Kang J. Analysis of traffic noise spatial distribution characteristics and influencing factors in high-density cities. Appl Acoustics. 2024;217:109838. 10.1016/j.apacoust.2023.109838Search in Google Scholar

[14] Poddar AK, Rohde UL, Apte AM. How low can they go?: Oscillator phase noise model, theoretical, experimental validation, and phase noise measurements. IEEE Microwave Magazine. 2013;14:50–72. 10.1109/MMM.2013.2269859Search in Google Scholar

[15] Givargis S, Karimi H. A basic neural traffic noise prediction model for tehran’s roads. J Environ Manag. 2010;91:2529–34. 10.1016/j.jenvman.2010.07.011Search in Google Scholar PubMed

[16] UKDT. 2024. Calculation of road traffic noise, UK department of transport. https://trid.trb.org/view/291939. Accessed on 2024-10-01. Search in Google Scholar

[17] Givargis S, Karimi H. Mathematical, statistical and neural models capable of predicting la, max for the Tehran-Karaj express train. Appl Acoustics. 2009;70:1015–20. 10.1016/j.apacoust.2008.11.003Search in Google Scholar

[18] Genaro N, Torija A, Ramos A, Requena I, Ruiz DP, Zamorano M. Modeling environmental noise using artificial neural networks. ISDA 2009: 2009 Ninth International Conference on Intelligent Systems Design and Applications. Pisa, Italy: IEEE; 2009. p. 215–9. 10.1109/ISDA.2009.179Search in Google Scholar

[19] Huang C, Zhang G, Yao J, Wang X, Calautit JK, Zhao C, et al. Accelerated environmental performance-driven urban design with generative adversarial network. Building Environ. 2022;224:109575. 10.1016/j.buildenv.2022.109575Search in Google Scholar

[20] Wang B, Kang J. Effects of urban morphology on the traffic noise distribution through noise mapping: A comparative study between UK and China. Appl Acoustics. 2011;72:556–68. 10.1016/j.apacoust.2011.01.011Search in Google Scholar

[21] Gharehchahi E, Hashemi H, Yunesian M, Samaei M, Azhdarpoor A, Oliaei M, et al. Geospatial analysis for environmental noise mapping: A land use regression approach in a metropolitan city. Environ Res. 2024;257:119375. 10.1016/j.envres.2024.119375Search in Google Scholar PubMed

[22] Han X, Huang X, Liang H, Ma S, Gong J. Analysis of the relationships between environmental noise and urban morphology. Environ Pollution. 2018;233:755–63. 10.1016/j.envpol.2017.10.126Search in Google Scholar PubMed

[23] Dempsey N, Brown C, Raman S, Porta S, Jenks M, Jones C, et al. Chromosome alterations in human solid tumors. In: Jenks M, Jones C, editors. Dimensions of the Sustainable City. Future City, Vol 2. Dordrecht: Springer Netherlands; 2010. p. 21–51. Search in Google Scholar

[24] Ryu H, Park IK, Chun BS, Chang SI. Spatial statistical analysis of the effects of urban form indicators on road-traffic noise exposure of a city in South Korea. Appl Acoustics. 2017;115:93–100. 10.1016/j.apacoust.2016.08.025Search in Google Scholar

[25] Silva LT, Oliveira M, Silva JF. Urban form indicators as proxy on the noise exposure of buildings. Appl Acoustics. 2014;76:366–76. 10.1016/j.apacoust.2013.07.027Search in Google Scholar

[26] Yin X, Fallah-Shorshani M, McConnell R, Fruin S, Franklin M. Predicting fine spatial scale traffic noise using mobile measurements and machine learning. Environ Sci Tech. 2020;54:12860–9. 10.1021/acs.est.0c01987Search in Google Scholar PubMed PubMed Central

[27] Fallah-Shorshani M, Yin X, McConnell R, Fruin S, Franklin M. Estimating traffic noise over a large urban area: An evaluation of methods. Environ Int. 2022;170:107583. 10.1016/j.envint.2022.107583Search in Google Scholar PubMed

[28] Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, et al. Generative adversarial networks. Commun ACM. 2020;63:139–44. 10.1145/3422622Search in Google Scholar

[29] Gonog L, Zhou Y. A review: generative adversarial networks. in: 2019 14th IEEE conference on industrial electronics and applications (ICIEA). IEEE; 2019. p. 505–10. 10.1109/ICIEA.2019.8833686Search in Google Scholar

[30] Wu AN, Stouffs R, Biljecki F. Generative adversarial networks in the built environment: A comprehensive review of the application of gans across data types and scales. Building Environ. 2022;223:109477. 10.1016/j.buildenv.2022.109477Search in Google Scholar

[31] Isola P, Zhu JY, Zhou T, Efros AA. Image-to-image translation with conditional adversarial networks. in: Proceedings of the IEEE conference on computer vision and pattern recognition; 2017. p. 1125–34. 10.1109/CVPR.2017.632Search in Google Scholar

[32] Zhu JY, Park T, Isola P, Efros AA. Unpaired image-to-image translation using cycle-consistent adversarial networks. in: Proceedings of the IEEE International Conference on Computer Vision; 2017. p. 2223–32. 10.1109/ICCV.2017.244Search in Google Scholar

[33] Liu Z, Yang J, He J, Li W, Qiu W. Gan-based transportation noise prediction via satellite maps: A case study in New York. J Digital Landscape Architecture. 2024;9:38–49. Search in Google Scholar

[34] Salomons EM, Pont MB. Urban traffic noise and the relation to urban density, form, and traffic elasticity. Landscape Urban Plan. 2012;108:2–16. 10.1016/j.landurbplan.2012.06.017Search in Google Scholar

[35] Mann S, Singh G. Traffic noise monitoring and modelling-an overview. Environ Sci Pollut Res. 2022;29:55568–89. 10.1007/s11356-022-21395-4Search in Google Scholar PubMed

[36] Garg N, Maji S. A critical review of principal traffic noise models: Strategies and implications. Environ Impact Assess Rev. 2014;46:68–81. 10.1016/j.eiar.2014.02.001Search in Google Scholar

[37] Staab J, Schady A, Weigand M, Lakes T, Taubenböck H. Predicting traffic noise using land-use regression–a scalable approach. J Exposure Sci Environ Epidemiol. 2022;32:232–43. 10.1038/s41370-021-00355-zSearch in Google Scholar PubMed PubMed Central

[38] Leitao AB, Ahern J. Applying landscape ecological concepts and metrics in sustainable landscape planning. Landscape Urban Plan. 2002;59:65–93. 10.1016/S0169-2046(02)00005-1Search in Google Scholar

[39] Attia S, Gratia E, De Herde A, Hensen JL. Simulation-based decision support tool for early stages of zero-energy building design. Energy Buildings. 2012;49:2–15. 10.1016/j.enbuild.2012.01.028Search in Google Scholar

[40] EPD, 2020 l10 (1 hour) at 4m. 2024. https://portal.csdi.gov.hk/. Accessed on 2024-10-01. Search in Google Scholar

[41] CSDI, 2024. https://portal.csdi.gov.hk/csdi-webpage/. Accessed on 2024-10-01. Search in Google Scholar

[42] EPD, 2020 l10 (1 hour) at 4m. 2024. https://portal.csdi.gov.hk/geoportal/?lang=en&datasetId=epd_rcd_1696985483224_32728. Accessed on 2024-10-01. Search in Google Scholar

[43] Chen J, Zheng H. Forecasting land surface drought in urban environments based on machine learning model. Sustainable Cities Soc. 2025;118:106048. 10.1016/j.scs.2024.106048Search in Google Scholar

[44] Pillow. 2025. https://python-pillow.org/. Accessed on 2025-04-08. Search in Google Scholar

[45] Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, et al. Generative adversarial nets. Adv Neural Inform Proces Syst. 2014;27. Search in Google Scholar

[46] Li Q, Zheng H. Prediction of summer daytime land surface temperature in urban environments based on machine learning. Sustain Cities Soc. 2023;97:104732. 10.1016/j.scs.2023.104732Search in Google Scholar

[47] Zhang Y, Zhang Q, Zhao Y, Deng Y, Zheng H. Urban spatial risk prediction and optimization analysis of poi based on deep learning from the perspective of an epidemic. Int J Appl Earth Observ Geoinform. 2022;112:102942. 10.1016/j.jag.2022.102942Search in Google Scholar PubMed PubMed Central

[48] Wang TC, Liu MY, Zhu JY, Tao A, Kautz J, Catanzaro B. High-resolution image synthesis and semantic manipulation with conditional GANs. in: Proceedings of the IEEE conference on computer vision and pattern recognition; 2018. p. 8798–807. 10.1109/CVPR.2018.00917Search in Google Scholar

[49] Salimans T, Goodfellow I, Zaremba W, Cheung V., Radford A, Chen X. Improved techniques for training gans. Advances in Neural Information Processing Systems 29 (NIPS 2016); 2016. p. 2234–42. Search in Google Scholar

[50] Heusel M, Ramsauer H, Unterthiner T, Nessler B, Hochreiter S. GANs trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems 30 (NIPS 2017). 2017; p. 6629–40.Search in Google Scholar

[51] Zhang R, Isola P, Efros AA, Shechtman E, Wang O. The unreasonable effectiveness of deep features as a perceptual metric. in: Proceedings of the IEEE conference on computer vision and pattern recognition. 2018. p. 586–95. 10.1109/CVPR.2018.00068Search in Google Scholar

[52] Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assessment: from error visibility to structural similarity. IEEE Trans Image Proces. 2004;13:600–12. 10.1109/TIP.2003.819861Search in Google Scholar

[53] LightningAI Torchmetrics. 2024. https://lightning.ai/docs/torchmetrics/stable/all-metrics.html. Accessed on 2024-10-01. Search in Google Scholar

[54] Zhu JJ, Yang M, Ren ZJ. Machine learning in environmental research: common pitfalls and best practices. Environ Sci Tech. 2023;57:17671–89. 10.1021/acs.est.3c00026Search in Google Scholar PubMed

[55] Margaritis E, Kang J. Relationship between green space-related morphology and noise pollution. Ecol Indicators. 2017;72:921–33. 10.1016/j.ecolind.2016.09.032Search in Google Scholar

[56] Margaritis E, Kang J. Relationship between urban green spaces and other features of urban morphology with traffic noise distribution. Urban Forestry Urban Green. 2016;15:174–85. 10.1016/j.ufug.2015.12.009Search in Google Scholar

[57] Ow LF, Ghosh S. Urban cities and road traffic noise: Reduction through vegetation. Appl Acoustics. 2017;120:15–20. 10.1016/j.apacoust.2017.01.007Search in Google Scholar

[58] Yuan M, Yin C, Sun Y, Chen W. Examining the associations between urban built environment and noise pollution in high-density high-rise urban areas: A case study in Wuhan, China. Sustain Cities Soc. 2019;50:101678. 10.1016/j.scs.2019.101678Search in Google Scholar

[59] Sandberg U. Road traffic noise-the influence of the road surface and its characterization. Appl Acoustics. 1987;21:97–118. 10.1016/0003-682X(87)90004-1Search in Google Scholar

[60] Tan JKA, Hasegawa Y, Lau SK, Tang SK. The effects of visual landscape and traffic type on soundscape perception in high-rise residential estates of an urban city. Appl Acoustics. 2022;189:108580. 10.1016/j.apacoust.2021.108580Search in Google Scholar

[61] Watts G, Crombie D, Hothersall D. Acoustic performance of new designs of traffic noise barriers: full scale tests. J Sound Vibration. 1994;177:289–305. 10.1006/jsvi.1994.1435Search in Google Scholar

[62] CSDI, 2024. Noise barrier. https://portal.csdi.gov.hk/geoportal/?lang=en&datasetId=hyd_rcd_1632360656336_40750. Accessed on 2024-10-01. Search in Google Scholar

[63] EPD, Mitigation measures. 2024. https://www.epd.gov.hk/epd/noise_education/web/ENG_EPD_HTML/m4/mitigation_2_more.html. Accessed on 2024-10-01. Search in Google Scholar

[64] CSDI. 2024. Noise enclosure. https://portal.csdi.gov.hk/geoportal/?lang=en&datasetId=hyd_rcd_1632360749014_39795. Accessed on 2024-10-01. Search in Google Scholar

[65] Cao R, Leng Z, Yu J, Hsu SC. Multi-objective optimization for maintaining low-noise pavement network system in Hong Kong. Trans Res D Transport Environ. 2020;88:102573. 10.1016/j.trd.2020.102573Search in Google Scholar

[66] Chui E, Lee C, Leung J, Yeung M. Applying low noise road surfacing to reduce road traffic noise in hong kong. ICA2016. 2016. Search in Google Scholar

[67] Ulf Sandberg. Review and evaluation of the low-noise road surface programme for low-speed roads in Hong Kong. 2008. https://www.epd.gov.hk/epd/sites/default/files/epd/english/environmentinhk/noise/studyrpts/files/LNRS-Mar08.pdf; https://www.epd.gov.hk/epd/sites/default/files/epd/english/environmentinhk/noise/studyrpts/files/LNRS-Mar08.pdf. Accessed on 2024-10-01. Search in Google Scholar

[68] Sakieh Y, Jaafari S, Ahmadi M, Danekar A. Green and calm: Modeling the relationships between noise pollution propagation and spatial patterns of urban structures and green covers. Urban Forestry Urban Greening. 2017;24:195–211. 10.1016/j.ufug.2017.04.008Search in Google Scholar

[69] Forssén J, Gustafson A, Pont MB, Haeger-Eugensson M, Achberger C, Rosholm N. Effects of urban morphology on traffic noise: A parameter study including indirect noise exposure and estimated health impact. Appl Acoustics. 2022;186:108436. 10.1016/j.apacoust.2021.108436Search in Google Scholar

[70] Tong H, Kang J. Relationship between urban development patterns and noise complaints in England. Environ Plan B Urban Analytics City Sci. 2021;48(6):1632–49. 10.1177/2399808320930247Search in Google Scholar

[71] Wang H, Sun B, Chen L. An optimization model for planning road networks that considers traffic noise impact. Appl Acoustics. 2022;192:108693. 10.1016/j.apacoust.2022.108693Search in Google Scholar

[72] Chen L, Sun B, Wang H, Li Q, Hu L, Chen Z. Forecast and control of traffic noise based on improved UE model during road network design. Appl Acoustics. 2020;170:107529. 10.1016/j.apacoust.2020.107529Search in Google Scholar

[73] Nega T, Smith C, Bethune J, Fuuu WH. An analysis of landscape penetration by road infrastructure and traffic noise. Comput Environ Urban Syst. 2012;36(3):245–56. 10.1016/j.compenvurbsys.2011.09.001Search in Google Scholar

[74] Calarco FM, Galbrun L. Sound mapping design of water features used over road traffic noise for improving the soundscape. Appl Acoustics. 2024;219:109947. 10.1016/j.apacoust.2024.109947Search in Google Scholar

[75] Bowlby W, Williamson R, Reiter D, et al. Field evaluation of reflected noise from a single noise barrier. Chater 6. 2018. p. 140. https://trid.trb.org/View/1566104. Accessed on 2025-09-30. Search in Google Scholar

[76] Day CW. 2005. Sound Reflection from a Motorway Barrier. Marshall Day Acoustics, maszynopis. https://www.acoustics.org.nz/sites/www.acoustics.org.nz/files/journal/pdfs/Day,_C_NZA2005.pdf. Accessed on 2025-09-30. Search in Google Scholar

[77] Zhang M, Wu Y, Li M, Meng Q, Kang J. Examining how green space patterns affect noise distribution in urban areas with varying built-up densities. J Environ Manag. 2025;393:127183. 10.1016/j.jenvman.2025.127183Search in Google Scholar PubMed

[78] UN. 2015. Sustainable development goals: 17 goals to transform our world https://www.un.org/en/exhibits/page/sdgs-17-goals-transform-world. Accessed on 2024-10-01. Search in Google Scholar

Received: 2025-04-26

Revised: 2025-10-16

Accepted: 2025-10-19

Published Online: 2025-11-24

This work is licensed under the Creative Commons Attribution 4.0 International License.