TL;DR
This paper introduces a method for creating statistical road-traffic noise maps using only basic urban form indicators, which is crucial for early-stage urban planning where detailed data is scarce. By developing and comparing Artificial Neural Network (ANN) and Ordinary Least Squares (OLS) models, the study demonstrates that road-related factors like traffic volume and road area density are key predictors of noise levels, offering a more accessible approach to noise mapping for sustainable city development.
What Problem Does This Paper Address?
Generating accurate road-traffic noise maps is essential for urban planning and mitigating noise pollution, which negatively impacts public health and quality of life. However, traditional noise mapping requires extensive data on buildings, roads, and traffic, along with sophisticated software and hardware. This makes it challenging to create noise maps during the early stages of urban planning when only rudimentary information like population density and general urban layout is available. The paper addresses this gap by proposing a statistical approach that leverages elementary urban form indicators to predict noise levels, thereby enabling noise mapping even in the absence of detailed urban morphology data.
Methods
The researchers developed two statistical models: an Artificial Neural Network (ANN) and an Ordinary Least Squares (OLS) model. Both models were trained using data from Gwangju, South Korea, which included urban form indicators derived from a 3D city model and GIS data, as well as road-traffic noise levels from an existing normal noise map.
- Urban Form Indicators: A comprehensive set of indicators related to population, buildings, roads, and land use were extracted. These included population density, ground space index (GSI), floor space index (FSI), traffic volume (Q), heavy vehicle ratio (PH), speed (V), traffic density (Dt), road area density (Ra), noise barrier area density (Wa), and fractions of residential (LR), commercial (Lc), industrial (LI), and green (LG) land use.
- Model Development:
- OLS Model: Used to identify statistically significant urban form indicators and to address multicollinearity among variables.
- ANN Model: Developed to capture complex, non-linear relationships between urban form indicators and noise levels. The ANN model's parameters (number of hidden nodes and decay parameter) were optimized to minimize the Root Mean Squared Error (RMSE) between predicted and actual noise levels.
- Validation: The developed ANN and OLS models were then applied to a different city, Cheongju, South Korea, to generate statistical noise maps. These maps were subsequently compared against the existing normal road-traffic noise map of Cheongju to evaluate the models' predictive accuracy and applicability.
Datasets
The study utilized data from two South Korean cities:
- Gwangju Metropolitan City: Served as the training and development dataset. Data included:
- 3D urban model and GIS data for urban form indicators.
- An existing normal road-traffic noise map (generated using SoundPLAN software).
- On-site measurements for verification.
- Cheongju: Served as the verification and testing dataset. Data included:
- An existing normal road-traffic noise map.
The paper also references various publicly available datasets for topography, LiDAR, road networks, and population data for both cities.
Key Insights and Ideas
- Accessibility of Noise Mapping: The core idea is that basic urban form indicators can be sufficient for generating useful statistical noise maps, making noise assessment more accessible for early-stage urban planning where detailed data is unavailable.
- Importance of Road-Related Factors: The study strongly highlights that road-related urban form indicators, specifically traffic volume and road area density, are the most significant predictors of road-traffic noise levels. This suggests that focusing on these factors can be highly effective in designing quieter urban environments.
- Model Performance Differences: While both ANN and OLS models were developed, they exhibited different prediction tendencies. The OLS model tended to under-estimate noise levels, whereas the ANN model tended to over-estimate them. This difference is crucial for understanding the biases of each model when interpreting the generated noise maps.
- ANN Optimization is Key: The paper demonstrates the importance of optimizing ANN parameters (like the number of hidden nodes and decay rate) to achieve better prediction accuracy, as shown by the RMSE investigation.
Key Results
- OLS Model Significance: In the OLS model, seven out of ten urban form indicators were found to be statistically significant predictors of road-traffic noise. These included FSI, Q, PH, V, Ra, LR, and Lc.
- ANN vs. OLS Accuracy: For Gwangju, the ANN model showed slightly better accuracy (correlation coefficient of 0.70, R-squared of 0.49) compared to the OLS model (correlation coefficient of 0.66, R-squared of 0.44) when compared to the normal noise map. However, for Cheongju, the OLS model performed slightly better (correlation coefficient of 0.70, R-squared of 0.49) than the ANN model (correlation coefficient of 0.66, R-squared of 0.44).
- Predictive Tendencies: Consistently across both cities, the OLS model tended to under-predict noise levels, while the ANN model tended to over-predict them.
- Most Important Predictors: The ANN model's importance analysis revealed that traffic volume (Q) was the most significant urban form indicator for predicting road-traffic noise. Road area density (Ra) was also identified as important.
Limitations and Future Work
- Grid Cell Resolution: A limitation of statistical noise maps is that they typically predict a single representative value for a grid cell. Unlike normal noise maps, which can offer variable resolutions, these statistical maps require a minimum area (like a grid cell) to extract urban form indicators.
- Spatial Dependency: The study acknowledges that it only considered urban form indicators within a single target grid cell. Future work could incorporate the spatial dependency of neighboring cells by employing models like Convolutional Neural Networks (CNNs), which are adept at image recognition and classification, to potentially improve prediction accuracy.
- Model Biases: The observed tendency of OLS to under-estimate and ANN to over-estimate noise levels suggests that users should be aware of these biases when interpreting the generated maps and making planning decisions. Further research could explore methods to mitigate these biases.
Summary updated