268 American Academic Scientific Research Journal for Engineering, Technology, and Sciences ISSN (Print) 2313-4410, ISSN (Online) 2313-4402 https://asrjetsjournal.org/index.php/American_Scientific_Journal/index A Smart Traffic Noise Prediction Model for Nairobi City, Kenya Ochungo Akech Elishaa*,Osano Simpson Nyambaneb ,Gichaga John Francisc aDepartment of Civil, Faculty of Engineering and Technology (FoET), Multimedia University, P.O. Box 15653- 00503 Nairobi, Kenya a,b,cDepartment of Civil and Construction Engineering, University of Nairobi, P.O. Box 30197-00100, Nairobi, Kenya aEmail:elishakech1@gmail.com,bEmail:sosano@uonbi.ac.ke,cEmail: fgichaga@uonbi.ac.ke Abstract Road traffic flow produces an undesirable externality since it distorts the ambient environmental noise, especially in cities. Such nuisance noise poses a risk to the health of the inhabitants. Globally, the combined concert of the forces of urbanization and road transport motorization has intensified the noise pollution challenge; yet, locally adapted predictive tools remain limited. In Nairobi, the capital city of Kenya, Road Traffic Noise (RTN) remains a less understood environmental nuisance. To date, no predictive RTN models have been developed, while established models such as CoRTN and RLS-90 lack applicability to Nairobi’s traffic and environmental conditions. This study aimed to develop an accurate smart model leveraging artificial neural networks (ANNs) to forecast RTN levels using traffic information data [22]. Traffic data, including audio recordings using a Samsung Galaxy A12 Model SM-A127F/DS Android Smartphone, equivalent noise levels (Leq) using a Lutron SL-4033SD Class 1 Sound Level Meter (SLM), vehicular volume using a manual tally form, and speed using a speed gun, was collected across 42 locations within Nairobi. Using this data, an Artificial Neural Network (ANN), Multi-Layer Perceptron (MLP) model, was developed with two hidden layers. Hyperparameter tuning via grid search was done to optimize model performance. The model achieved a Mean Absolute Error (MAE) of 0.97 dBA and an R2 value of 0.90, outperforming traditional statistical models like CoRTN with an MAE of 5.0 dBA and RLS-90 with an MAE of 11.0 dBA. These results highlight the model’s high accuracy in predicting Nairobi’s RTN. The model’s deployment on a web-based dashboard enables real-time noise monitoring and stakeholder engagement. This pioneering smart predictive model for Nairobi offers a scalable solution for urban noise management [25], with potential applications in traffic planning and policy implementation. Keywords: Road Traffic Noise; Machine Learning; Smart prediction model. ------------------------------------------------------------------------ Received: 8/15/2025 Accepted: 10/15/2025 Published: 10/25/2026 ------------------------------------------------------------------------ * Corresponding author. https://asrjetsjournal.org/index.php/American_Scientific_Journal/index American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 269 1. Introduction Road traffic noise (RTN) is a pressing environmental and public health concern in urban areas, particularly in rapidly urbanizing cities worldwide. As a byproduct of vehicular traffic, RTN contributes to adverse health outcomes, including stress, sleep disturbances, and cardiovascular diseases [28], as documented by the World Health Organization [1]. Studies by [2,3] highlight noise’s auditory and non-auditory effects, such as anxiety and cognitive impairment, necessitating robust noise management. In Europe and North America, long-term monitoring has linked RTN to reduced well-being, with studies like [4,5,6,20] showing increased mortality, mental health issues, and cardiovascular risks.In developing regions like Asia and Africa, dense traffic and heterogeneous vehicle compositions exacerbate RTN challenges [26]. For instance, studies in Delhi, India, have reported noise levels consistently exceeding permissible limits across major traffic corridors [7,8]. In Africa, research in Lagos, Nigeria, documented chronic noise exposure surpassing WHO guidelines, posing significant health risks [9]. Similarly, [10] applied spatial modeling to map environmental noise in Accra, Ghana, revealing exposure inequalities in urban settings. Earlier work by [11] further confirmed that traffic-related noise contributes to cardiovascular disease risk, highlighting the global relevance of RTN as a public health issue.While measurement- based studies provide critical insights, their high costs and logistical complexity make them impractical for large- scale urban applications, driving demand for predictive models tailored to local conditions [29]. Traditional statistical models, such as CoRTN and RLS-90, rely on traffic volume, speed, and road conditions but often underperform in non-European contexts due to differing traffic patterns [12]. Recent advancements in machine learning, particularly Artificial Neural Networks (ANNs) and ensemble methods, have demonstrated superior performance in capturing complex, non-linear relationships in traffic noise data. For example, [13] developed an emotional ANN model for vehicular traffic noise in Tehran, achieving high accuracy by modeling non-linear traffic dynamics. Similarly, [14] employed an ANN with grid search hyperparameter tuning to predict RTN in Bogotá, optimizing for urban-specific variables. [15] applied ensemble methods, such as random forests, to predict road traffic noise, emphasizing their robustness in handling diverse traffic datasets. [16] leveraged deep learning with convolutional layers to forecast urban noise in real time, incorporating spatial-temporal features. [17] used ANNs to model highway noise in India, while [18] integrated ANN with contouring techniques for spatial noise mapping. [19] combined ensemble machine learning with GIS to predict campus traffic noise, highlighting the role of spatial variables in enhancing model accuracy. These studies collectively demonstrate that machine learning approaches, when adapted to local conditions, outperform traditional statistical models by leveraging advanced mathematical frameworks to model complex urban noise patterns. In Eastern Africa, smart predictive models for RTN remain scarce, with most studies relying on outdated statistical methods. Nairobi, Kenya’s capital, presents unique challenges due to its diverse vehicle fleet, including bicycles, motorcycles, cars, buses, and trucks, operating within constrained road networks (Appendix 4). This heterogeneity, coupled with frequent congestion, results in irregular traffic flow and elevated noise levels. Existing models like CoRTN and RLS-90 are not calibrated for Nairobi’s conditions, highlighting a critical gap in localized RTN prediction frameworks. This study addresses this gap by developing the first ANN-based RTN prediction model tailored to Nairobi’s traffic and environmental context. Its novelty lies in using a Multi-Layer Perceptron (MLP) ANN, optimized for Nairobi’s heterogeneous traffic patterns via grid search hyperparameter tuning, and deploying it on a web-based dashboard for real-time monitoring for noise reduction schemes [23]. The study American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 270 contributes: (1) a high-accuracy MLP model for RTN prediction, (2) a scalable framework for noise management in African cities, and (3) a public platform for stakeholder engagement and urban noise policy formulation as suggested by [24]. These advancements aim to support urban highway planning [27], mitigate health impacts, and enhance liveability in Nairobi and similar contexts. The source code for the MLP ANN model, including data preprocessing, model training, hyperparameter tuning, and web-based dashboard deployment, is available on GitHub at https://github.com/ElishaAkech/SOUNDAI. This repository includes Python scripts for implementing the model using libraries such as PyTorch, along with real datasets and instructions for reproducing the results. 2. Methodology 2.1 . Study Area The study was conducted in Nairobi City, Kenya, a bustling metropolis with a diverse vehicle fleet. Figure 1: Map of Nairobi City, Kenya (Source: Author) 42 sampling points were selected across the city, spanning diverse land use types connected by diverse road corridors that have different traffic volumes, all representing varied urban traffic conditions, see Figure 2. American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 271 Figure 2: 42 Sampling locations selected across Nairobi (Source: Author) 2.2 . Data Collection Data were collected for seven days, across 42 Nairobi locations from 6 AM to 6 PM. Noise levels (Leq) were measured using a Lutron SL-4033SD Class 1 calibrated Sound Level Meter (SLM). A handheld Samsung Galaxy A12 Model SM-A127F/DS, Android smartphone was also used to capture audio recordings of RTN. Vehicle count spanning bicycles, motorcycles, private cars, SUVs, Pick-ups, Public Service Vehicles (PSVs), buses, light, medium, and heavy-duty trucks, and others such as tractors were manually tallied on a form shown in Appendix 1. Traffic speed was also captured using a calibrated Binar Radar speed gun and documented manually on a form shown in Appendix 1. From the data, a total of 504 samples of data were obtained. 2.3 . Data Preprocessing Noise levels were calculated in MS Excel from the logged SLM data. Vehicle counts were converted to Passenger Car Units (PCU) using standard conversion factors shown in Appendix 3. Average speeds were also calculated from the speed gun data, and the flow type was categorized as congested (<20 km/h), periodic (20–35 km/h), or fluid (>35 km/h). The smartphone audio recordings were preprocessed using Python to extract the Leq. They were compared with the SLM measurements to ensure consistency. 2.4 . Framework Overview The modelling process followed a structured pipeline, as shown in Figure 3, from data sources to evaluation, with shared preprocessing and diverse modelling approaches. The framework as presented in Figure 3, includes: (a) American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 272 Data Sources: Three Excel sheets with the Noise Leq, Speed, and Vehicle Counts in PCU, per location and at different hour bands, providing the raw data, (b) Common Preprocessing: Shared steps across all models, including feature en- gineering (motorcycles, light/medium/heavy vehicles, speed, lanes, flow type), data processing and cleaning, and an 80/20 train/test split, (c) Model Section: Divided into three categories: Traditional Machine Learning Models, that is, Random Forest, XGBoost, SVR, Custom ANN that is the 3-layer network implemented in this study, and Research-Based ANNs, that is, models from literature (Cammarata, Bogotá, Tehran, UAE, Genaro, and Torija), and Evaluation: Common performance metrics (Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the coefficient of determination (R2)) for comparing all models. This pipeline emphasizes that while preprocessing is identical, providing a standardized feature set to all models, the modelling approaches differ, utilizing distinct architectures and algorithms to predict the Leq. Figure 3: Visual pipeline diagram illustrating the modelling framework (Source: Author) American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 273 2.5. Smart Prediction Model A Multi-Layer Perceptron (MLP) ANN was selected due to its superior handling of non-linear relationships, after comparing different algorithms such as Random Forest, XGBoost, SVR, and Linear Regression, see Figure 3 above. 2.6. Steps of Modelling 2.6.1. Input Vector Definition The input to the neural network is formalized as a vector x, each component representing a key traffic or environmental feature contributing to RTN. x is defined as: x = [x1, x2, x3, x4, x5, x6, x7, x8]𝑇 The superscript T means transpose, that is, turning a row into a column for math purposes. x1 denotes the count of motor cycles, the high-frequency noise contributors; x2 is the count of light vehicles, that is, cars, which are the primary volume drivers; x3 medium vehicles that is, vans, which produce moderate noise; x4 represents heavy vehicles, that is, trucks, which are low-frequency dominant; x5 is average vehicle speed in kilometers per hour, influencing Doppler effects and tire-road interactions; x6 represents the number of lanes, affecting noise propagation; x7 represents the passenger car units (PCU), which is a standardized measure of traffic volume, and x8 represents the flow type (categorical: 0 for congested, 1 for periodic, 2 for fluid), impacting noise variability. These features were selected based on correlation analysis and domain knowledge to capture the stochastic nature of urban RTN in Nairobi. 2.6.2. Forward Propagation The network processes the input through a series of linear transformations and non-linear activations to model complex RTN patterns. The computations are: The first hidden layer: h1 = max (0, W1x + b1) Equation 1 Where W1 ∈ R25×8 represents the weight matrix. It is a table of numbers that adjusts how much each input affects the layer. R25×8 means 25 rows by 8 columns of real numbers. x is the input vector, and b1 ∈ R25 is the bias vector, which adds a constant to shift the output. max (0,) is the Rectified Linear Unit (ReLU) activation function, which introduces non-linearity (allows modeling curves, not just straight lines) to handle heteroscedastic noise data (varying error) and prevents vanishing gradients (a training problem where updates become too small) The second hidden layer: h2 = max (0, W2h1 + b2) Equation 2 Where W2 ∈ R50×25, b2 ∈ R50 represents the weight matrix and the bias vector that expand feature representations American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 274 for deeper pattern recognition. The output layer (the Nairobi’s Smart RTN Prediction Model): y = W3h2 + b3 Equation 3 Where W3 ∈ R1×50, and b3 ∈ R, yielding the predicted equivalent noise level, y, in dBA. This architecture allows the model to learn hierarchical features, from raw traffic counts to aggregated noise predictions, optimizing for the non-stationary characteristics of RTN. 2.6.3. Data Splitting and Cross-Validation The dataset of 504 samples was partitioned into a training set (80%, 404 samples) and a testing set (20%, 100 samples) using stratified sampling to maintain distribution balance across noise hotspots. This split ensures robust generalization while allocating sufficient data for learning. To mitigate overfitting and assess model stability, 5- fold cross-validation was employed on the training set: the data is divided into 5 subsets, with each fold used once as validation while training on the remaining four. This process yields averaged performance metrics, providing a scientifically rigorous estimate of out-of-sample errors in the context of variable Nairobi traffic conditions. 2.6.4. Hyperparameter Optimization and Evaluation Hyperparameters, including learning rate (range: [0.0001, 0.01]), batch size (range: [16, 64]), and number of epochs (up to 500 with early stopping), were tuned using grid search, exhaustively evaluating combinations to minimize validation loss. Model performance was quantified using: Mean Absolute Error (MAE): 𝑀𝐴𝐸 = 1 𝑁 ∑ |(𝑦𝑖 − 𝑦�̂�| 𝑁 𝑖=1 Equation 4 Where N is the number of samples, for example, 100 in the test set, yi is the observed Leq at sample i, and 𝑦�̂� is the predicted Leq at sample i. MAE measures average deviation in dBA and is crucial for practical noise forecasting. Root Mean Squared Error (RMSE): RMSE = √ 1 𝑁 ∑ (𝑦𝑖 − 𝑦�̂�) 2𝑁 𝑖=1 Equation 5 Where N is the number of samples, yi is the observed Leq at sample i, and 𝑦�̂� is the predicted Leq at sample i. The RMSE emphasizes larger errors in high-noise scenarios. The Coefficient of Determination (R2): 𝑅2 = 1 − ∑(𝑦𝑖−𝑦�̂�)2 ∑(𝑦𝑖−�̅�)2 Equation 6 Where �̅� is the mean of the observed Leqs, yi is the observed Leq at sample i, and 𝑦�̂� is the predicted Leq at sample i. The coefficient of determination indicates an explained variance; a value closer to 1 is better, meaning that the model will account for most differences in the data. American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 275 Pearson correlation coefficient: 𝑟 = ∑(𝑦𝑖−�̅�)(𝑦𝑖−�̅̂�) √∑(𝑦𝑖−�̅�)2 ∑(𝑦𝑖−�̅̂�)2 Equation 7 Where yi is the observed Leq at sample i, 𝑦�̂� is the predicted Leq at sample i, �̅� is the mean of the observed Leqs, and �̅̂� is the mean of predicted Leqs. It assesses linear agreement between predicted and actual Leq, ensuring scientific validity. 2.6.5. Loss Function The Mean Squared Error (MSE) was selected as the objective function for optimization, where: 𝑀𝑆𝐸 = 1 𝑁 ∑ (𝑦𝑖 − 𝑦�̂�) 2𝑁 𝑖=1 Equation 8 Where N is the number of samples, yi is the observed Leq at sample i, and 𝑦�̂� Is the predicted Leq at sample i. The MSE is quadratic, penalizing larger deviations more severely, which is appropriate for regression tasks like RTN prediction, where minimizing variance in noise estimates is critical for public health applications. It aligns with the Gaussian assumption of noise residuals in environmental modeling. 2.6.6. Optimizer The Adam (Adaptive Moment Estimation) optimizer was utilized for efficient gradient descent, updating the weights as: 𝑤𝑡+1 = 𝑤𝑡 − 𝜂 ∙ 𝑚�̂� √𝑣�̂�+𝜖 Equation 9 where wt is the weight at step t (current), wt+1 is the updated weight, η is the learning rate (step size), �̂�𝑡 is the bias-corrected first moment estimate (smoothed gradient, like momentum),𝑣�̂� is the bias-corrected second moment estimate (smoothed squared gradient for adaptive rates), and 𝜖 = 10−8, for numerical stability (prevents division by zero). Adam combines momentum and RMSprop (adapts rates per parameter) advantages, adapting per-parameter learning rates, which accelerates convergence on the non-convex loss landscape of ANN training for RTN data with inherent multicollinearity. American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 276 Figure 4: MLP Architecture for RTN Prediction 2.7. Detailed Explanation of Forward Propagation and Optimization 2.7.1. Forward Propagation Components Forward propagation is the process by which the input data flows through the network layers to produce an output prediction. Imagine it as a factory assembly line where raw materials (inputs) are transformed step by step. The key variables are:  Weight Matrices (W): These are like adjustable knobs in the network. The weight matrices W1 ∈ R25×8 (25 by 8 grid of numbers), W2 ∈ R50×25, and W3 ∈ R1×50 contain learnable parameters that connect neurons (processing units) between layers. For instance, W1 weights the 8 input features, for example, motorcycle counts as x1, and speed as x5, to the 25 neurons in the first hidden layer, scaling each feature’s contribution to capture its influence on noise. These weights are changed during training to make better predictions. A high weight on x1 might mean motorcycles are very important for noise in Nairobi.  Hidden Layer Outputs (h): These are intermediate results. The vectors h1 ∈ R25 (list of 25 numbers) and h2 ∈ R50 represent the activations (outputs) of the first and second hidden layers, respectively. Computed as in Equation 1, multiply weights by inputs, add bias, then set negatives to zero. Similarly, for h2 as in Equation 2, they apply the ReLU activation to introduce non-linearity, enabling the model to learn complex patterns, like how motorcycle counts and traffic flow together affect noise.  Bias Vectors (b): These are constants added to adjust the output. The bias terms b1 ∈ R25 (25 numbers), b2 ∈ R50, and b3 ∈ R (single number) shift the linear transformations in each layer, allowing the network to fit data better. For example, b3 adjusts the final prediction �̂� to account for baseline noise, even if all inputs are American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 277 zero.  Observed Leq (yi): This is the real, measured noise level (Leq, in dBA) for the i-th data sample (where i goes from 1 to N, the total number of samples). It is collected from actual measurements taken and is the "correct answer" the model tries to match.  Predicted Leq (𝑦�̂�): This is the model’s guess for the noise level for the i-th sample, calculated as in Equation 3. It is compared to yi to see how wrong the model is, and is shown on the dashboard for users. 2.7.2. Adam Optimizer Components The optimizer is like a teacher that corrects the model’s mistakes by adjusting weights. The Adam optimizer updates the weights and biases to make the loss (error) smaller, using Equation 8. Think of it as taking small steps downhill to find the lowest error. The components are:  wt and wt+1: A single weight (one number in W) at the current step t (like time step in training), and its new value after update. This happens for every weight and bias to improve the model.  Learning Rate (η): This is how big each step is (tuned between 0.0001 and 0.01). Too big, and you might overshoot; too small, and learning is slow. It is like the stride length when walking downhill.  Bias-Corrected First Moment (𝑚�̂�): This is a smoothed version of the gradient (direction of steepest descent, calculated as change in loss (𝜕𝐿) per change in weight (𝜕𝑊): 𝑔𝑡 = 𝜕𝐿 𝜕𝑊 Equation 10 𝑚�̂� = 𝑚𝑡 1−𝛽𝑡 Equation 11 Where: 𝑚𝑡 = 𝛽1𝑚𝑡−1 + (1 − 𝛽1)𝑔𝑡 Equation 12 and β1 = 0.9 (a decay factor). It adds momentum, like pushing a ball to keep it rolling in the same direction.  Bias-Corrected Second Moment (𝑣�̂�): This smoothens the squared gradient for adaptive steps: 𝑣�̂� = 𝑣𝑡 1−𝛽2 𝑡 Equation 13 with 𝑣𝑡 = 𝛽2𝑣𝑡−1 + (1 − 𝛽2)𝑔𝑡 2 Equation 14 American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 278 and β2 = 0.999. It makes steps smaller for noisy directions (high variance) and larger for consistent ones.  Numerical Stability Term (ϵ): A tiny number (10−8) added to avoid dividing by zero if 𝑣�̂� is very small, thus keeping calculations safe. This setup helps the model learn efficiently from Nairobi’s traffic data, handling complexities like varying vehicle noise. 2.8. Statistical Descriptors for Input Data Table in Appendix 4 presents statistical descriptors for the input data with a significance value α = 5% for the Kolmogorov-Smirnov (K-S) test. Statistical descriptors summarize data. �̅� is the mean (average value), σ is the standard deviation (how it is spread out), Min and Max are the smallest and largest, Range is the Maximum minus the Minimum, IQR is the interquartile range (middle 50% spread), C.V. is the coefficient of variation (relative spread in %), Kurtosis measures tail heaviness (high means more extremes), Asy. Coe. is asymmetry (skewness, positive means tail to the right), Kol. Smi. is the K-S test p-value (low means not normal distribution), Proportion is % of vehicle types. 2 . 8 . 1 . E v a l u a t i o n Performance was evaluated using MAE, RMSE, and R2, with Pearson correlation analysis between predicted and actual Leq, obtained from SLM data. The performance metrics were compared to those of traditional machine learning models and literature-based models. 2.8.2. M o d e l D e p l o y m e n t The model is deployed on a web-based dashboard, accessible to the public. Users input parameters such as the location, time, speed, and vehicle count, and the dashboard outputs predicted Leq in real-time, visualized via interactive charts. The platform supports noise monitoring, public education, and integration with traffic management systems. 3.Results The table in Appendix 5 shows the equivalent sound levels measured at different time intervals using the SLM for the 42 sampling locations 3.1. Nairobi Road Traffic Noise Prediction Model The Nairobi’s Smart RTN Prediction Model (MLP) 𝐋𝐞𝐪𝐩𝐫𝐞𝐝𝐢𝐜𝐭𝐞𝐝 = 𝐖𝟑𝐡𝟐 + 𝐛𝟑 Equation 15 Where h2 is defined as in Equation 2 and W3 ∈ R1×50, and b3 ∈ R represents the weight matrix and bias for the output layer (single prediction). American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 279 3.2. Model evaluation and validation The MLP model achieved an MAE of 0.97 dBA, RMSE of 1.38 dBA, and R2 of 0.90 with a Pearson Correlation coefficient of 0.9476 between the predicted and the measured Leq values, indicating strong predictive accuracy. Table 2 compares the MLP with other models, showing superior performance over CoRTN, which has an MAE of 5.0 dBA and an R² of 0.80, and RLS-90, which has an MAE of 11.0 dBA and an R² of 0.50, attributed to its ability to capture Nairobi’s unique traffic patterns. The UAE ANN (Mansourkhaki, Berangi, Haghiri, and Constantinescu, 2018) and Torija ANN (Torija and his colleagues, 2012) performed similarly well; however, the Current MLP is optimized for Nairobi’s context. Table 1: Model evaluation and validation with existing predictive models Model MAE (dBA) RMSE (dBA) R2 Current MLP 0.97 1.38 0.90 Bogotá MLP [14] [14] 1.19 1.56 0.87 ANN (Torija) [8] [8] 0.87 1.08 0.94 ANN (UAE) [11] [11] 0.79 0.97 0.95 XGBoost 1.14 1.39 0.89 SVR 2.67 3.59 0.30 Random Forest 1.09 1.43 0.89 Linear Regression 3.74 4.44 -0.07 CoRTN 5.00 – 0.80 RLS-90 11.00 – 0.50 Figure 5: Average Difference Between Measured and Predicted Leq American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 280 Figure 6: An example of a day SPL variation prediction for Location 1 The plot in Figure 6 above shows the actual and predicted change in Leq over time, illustrating the differences between measured and model-predicted noise levels per hourly interval from 6 AM to 6 PM. The close alignment, especially during peak hours, demonstrates the model’s capability to capture temporal variations in RTN, with minor deviations reflecting real-world complexities like unmodeled variables. 4.Discussion From the results of this study, it can be deduced that, the MLP model developed effectively captures Nairobi’ city’s complex traffic dynamics, including the high prevalence of motorcycles and variable flow types, which contribute significantly to road traffic noise (RTN) variability. As shown in Table 1, the model outperforms traditional models like CoRTN and RLS-90, primarily due to its ability to adapt to Nairobi’s heterogeneous vehicle fleet and congested road networks, which these conventional models fail to address. This adaptability stems from the model’s use of a Multi-Layer Perceptron (MLP) ANN, optimized through grid search hyperparameter tuning to handle non-linear relationships in traffic data. The model’s predictive accuracy supports cost-effective noise monitoring, offering a scalable solution for urban planning in rapidly growing African cities like Nairobi. By deploying the model on a web-based dashboard, it enables real-time noise prediction, facilitates stakeholder engagement, and supports public education on noise pollution. The platform’s integration with traffic management systems can inform urban noise policies and mitigate health impacts, such as stress and subjective annoyance. However, the model’s reliance on data from this study limits its ability to account for seasonal traffic variations or unmodeled variables like road surface type and weather conditions, which future research should address to enhance robustness. American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 281 Figure 7: Performance Comparison (All Models) bar chart showing MAE, RMSE, and R2 for each model Figure 8: Box Plot of Prediction Errors for All Models Predictive modeling offers significant advantages, including cost-effective noise monitoring and scalability, which are particularly beneficial for urban planning in rapidly growing cities like Nairobi. The deployment of the model on a web-based dashboard facilitates real-time noise prediction [21], enables stakeholder engagement, allows for public education on noise pollution, and integrates with traffic management systems to mitigate noise at major hotspots. Practical implications include informing urban noise policies and reducing health impacts like stress and subjective annoyance. However, its limitations lie in the model’s reliance on this study’s data, which may not account for seasonal traffic variations, and the absence of variables like road surface type or weather conditions. 5.Conclusion This study developed the first smart RTN prediction model for Nairobi, leveraging an MLP ANN with high American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 282 accuracy and surpassing traditional models like CoRTN and RLS-90. The model, tailored to Nairobi’s traffic dynamics, is deployed on a public web dashboard, enabling real-time noise monitoring and prediction and citizen engagement. The study recommends including real-time data integration, expanding input variables such as road surface and weather, and collaboration with traffic authorities to enhance urban noise management. This pioneering model sets a baseline for smart noise prediction in African cities, with the potential for broader application. References [1] World Health Organization, 2018. Environmental Noise Guidelines for the European Region. World Health Organization. [2] Begam, Nazia & Bashar, Abu. (2020). Auditory effects and consequences of noise pollution in humans: A scoping review. Journal of Family Medicine & Community Health. [3] Stansfeld SA, and Matheson MP (2003). Noise pollution: non-auditory effects on health. Br Med Bull. 2003; 68:243-57. doi: 10.1093/bmb/ldg033. PMID: 14757721. [4] Pascale Antonio, Claudio Guarnaccia, Eloísa Macedo, Paulo Fernandes, Ana I. Miranda, Susana Sargento, Margarida C. Coelho, (2023). Road traffic noise monitoring in a Smart City: Sensor and Model-Based approach, Transportation Research Part D: Transport and Environment, Volume 125,2023,103979, ISSN 1361-9209, https://doi.org/10.1016/j.trd.2023.103979. [5] Ranpise, Ramesh B. and Tandel, Bhaven N.(2022).Urban road traffic noise monitoring, mapping, modelling, and mitigation: A thematic review" Noise Mapping, vol. 9, no. 1, 2022, pp. 48-66. https://doi.org/10.1515/noise-2022-0004 [6] Ziqin Lan, and Ming Cai,(2021).Dynamic traffic noise maps based on noise monitoring and traffic speed data,Transportation Research Part D: Transport and Environment,Volume 94,2021,102796,ISSN 1361- 9209,https://doi.org/10.1016/j.trd.2021.102796. [7] Garg N. and Maji S.( 2016). A retrospective view of noise pollution control policy in India: status, proposed revisions and control measures. Current Science , 10 July 2016, Vol. 111, No. 1 (10 July 2016), pp. 29-38 [8] Narendra, S.; Davar, S.C (2004). Noise pollution-sources, effects and control. J. Hum. Ecol. 2004, 16, 181– [9] Atilade, A., Coker, J., Albert, I., & Ogede, R. (2020). Evaluation and analysis of traffic noise in Lagos State Polytechnic, Ikorodu, Lagos State, Nigeria. Nigerian Journal of Physics, 28, 179–210. [10] Clark, S. N., Alli, A. S., Ezzati, M., Brauer, M., Toledano, M. B., Nimo, J., Moses, J. B., Baah, S., Hughes, A., Cavanaugh, A., Agyei-Mensah, S., Owusu, G., Robinson, B., Baumgartner, J., Bennett, J. E., & Arku, R. E. (2022). Spatial modelling and inequalities of environmental noise in Accra, Ghana. Environmental Research, 214(Pt 2), 113932. https://doi.org/10.1016/j.envres.2022.113932 [11] Tétreault LF, Perron S, Smargiassi A. Cardiovascular health, traffic-related air pollution and noise: are associations mutually confounded? A systematic review. Int J Public Health. 2013 Oct;58(5):649-66. doi: 10.1007/s00038-013-0489-7. Epub 2013 Jul 26. PMID: 23887610; PMCID: PMC3907786. [12] Zhang, Y., Zhang, X., & Wang, Y. (2024). Urban traffic noise prediction using hybrid machine learning models. Journal of Environmental Management, 350, 119678. [13] Nourani, V., Gökçekus¸, H., Umar, I.K., 2020. Artificial intelligence-based modeling for predicting the https://doi.org/10.1016/j.trd.2023.103979 https://doi.org/10.1515/noise-2022-0004 https://doi.org/10.1016/j.envres.2022.113932 American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 283 road traffic noise. Appl. Acoust. 157, 107000. [14] Acosta, Ó., Montenegro, C., González Crespo, R., 2024. Road traffic noise prediction model based on artificial neural networks. Heliyon 10, e36484. https://doi.org/10.1016/j.heliyon.2024.e36484. [15] Sharma, A., Vijay, R., & Sohony, R. A. (2022). Application of machine learning techniques for predicting road traffic noise. Environmental Monitoring and Assessment, 194(3), 189. [16] Lee, J., Lee, J., & Kim, J. (2023). Deep learning-based traffic noise prediction in urban environments. Applied Acoustics, 201, 109090. [17] Kumar, P., Nigam, S. P., & Kumar, N. (2021). Vehicular traffic noise modeling using artificial neural network approach. Transportation Research Part D: Transport and Environment, 92, 102716. [18] Debnath, A., Singh, P. K., & Banerjee, S. (2022). Vehicular traffic noise modelling of urban area—a contouring and artificial neural network-based approach. Environmental Science and Pollution Research, 29(26), 39948–39972. [19] Umar, I. K., Nourani, V., & Gökçekuş, H. (2023). Traffic noise prediction model using GIS and ensemble machine learning: a case study at Universiti Teknologi Malaysia (UTM) Campus. Environmental Science and Pollution Research [20] European Environment Agency, 2020. Environmental Noise in Europe. European Environment Agency. [21] Li, H., & Zhang, J. (2023). Real-time traffic noise prediction using deep neural networks in smart cities. Sustainable Cities and Society, 89, 104354 [22] Cammarata, G., Cavalieri, S., Fichera, A., 1995. A neural network architecture for noise prediction. Neural Netw. 8, 963–973. [23] Genaro, N., Torija, A., Ramos-Ridao, A., Requena, I., Ruiz, D.P., Zamorano, M., 2010. Modeling of acoustic properties in the grasslands and noise reduction by hedges. Appl. Acoust. 71, 1065–1074. [24] Hamad, K., Khalil, M.A., Shanour, A.S., 2017. Modeling and analysis of traffic noise level using artificial neural network. A case study in Menoufia University, Egypt. Egyptian J. Ear Nose Throat Allied Sci. 18, 27–32. [25] Torija, A.J., Ruiz, D.P., Ramos-Ridao, A.F., 2012. Application of a methodology for categorizing and differentiating urban soundscapes using acoustical descriptors and semantic- differential attributes. J. Acoust. Soc. Am. 131, 3228–3231. [26] Mansourkhaki, A., Berangi, M., Haghiri, M., Constantinescu, M., 2018. Road traffic noise prediction model based on the traffic composition. J. Acoust. Soc. Am. 143, 3022–3031. [27] Kumar, K., Parida, M., Katiyar, V.K., 2013. Optimized height of noise barrier for non- urban highway using artificial neural network. Int. J. Environ. Sci. Technol. 11, 719–730. [28] Cai, Y., Hodgson, S., Blangiardo, M., Gulliver, J., Morley, D., Fecht, D., Vienneau, D., de Hoogh, K., Key, T., Hveem, K., Elliott, P., & Hansell, A. L. (2018). Road traffic noise, air pollution and incident cardiovascular disease: A joint analysis of the HUNT, EPIC-Oxford and UK Biobank cohorts. Environment international, 114, 191–201. https://doi.org/10.1016/j.envint.2018.02.048 [29] Tashakor, S., Chamani, A., & Moshtaghie, M. (2023). Noise pollution prediction and seasonal comparison in urban parks using a coupled GIS-artificial neural network model. Environmental Monitoring and Assessment, 195(2). https://doi.org/10.1016/j.heliyon.2024.e36484 https://doi.org/10.1016/j.envint.2018.02.048 American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 284 6.Appendix 1 Manual Tally form used for traffic count Figure 9 American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 285 Figure 10 Appendix 2 Manual Tally form used for recording speed Figure 11 American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 286 Figure 12 Appendix 3 Table2: PCU conversion factors Bicycle Motorcycle Private car Pickup SUV PSVs Buses Light trucks Medium trucks Heavy trucks Others 0.5 1 1 1 1 1.5 4 1.5 5 8 8 American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 287 Appendix 4 Table 3: showing the Statistical descriptors for input data (significance value α = 5% for K-S) Stat. Flow Motorcycles Light Medium Heavy q (PCU) Speed Lanes x̄ 11.65 58.24 12.83 5.71 1838.07 50.19 3.07 0.87 σ 15.23 38.76 7.45 6.12 580.34 14.82 0.98 0.34 Min. 0.00 5.00 0.00 0.00 409.00 25.00 2.00 0.00 Max. 63.00 210.00 42.00 21.00 6965.00 85.21 4.00 1.00 Range 63.00 205.00 42.00 21.00 6556.00 60.21 2.00 1.00 IQR 20.00 50.00 15.00 8.00 1200.00 20.00 2.00 1.00 C.V. (%) 130.69 66.55 58.06 107.18 31.57 29.53 31.92 39.08 Kurtosis 5.12 4.89 3.45 4.23 3.12 2.89 1.45 1.23 Asy. Coe. 2.34 2.10 1.87 2.01 1.65 1.23 0.67 0.45 Kol. Smi. p < 0.001 p < 0.01 p < 0.05 p < 0.01 p < 0.001 p < 0.05 p < 0.1 p < 0.05 Proportion 100.00% 24.37% 49.84% 14.84% 9.94% 100.00% 100.00% 100.00% American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 288 Appendix 5 Table 4: showing the measured/observed RTN levels in Nairobi, Kenya LOCATION 6 AM-7 AM 7 AM-8 AM 8 AM-9 AM 9 AM-10 AM 10 AM-11 AM 11 AM- 12 PM 12 PM-1 PM 1 PM-2 PM 2 PM-3 PM 3 PM-4 PM 4 PM-5 PM 5 PM-6 PM 1 73.52 72.12 71.32 71.92 73.02 74.32 74.92 74.22 72.02 71.32 71.12 71.02 2 82.89 81.39 80.69 81.29 82.39 83.69 84.29 83.29 81.39 80.59 80.39 80.29 3 75.99 74.39 73.79 74.39 75.49 76.79 77.39 76.29 75.09 74.39 73.59 73.39 4 72.2 70.8 71 71.6 72.7 74 74.6 73.5 71.7 71 70.7 70.6 5 75.35 74.45 74.85 75.45 75.95 77.85 78.45 77.35 75.55 74.85 74.65 74.55 6 82.92 81.22 81.72 82.32 83.42 84.72 85.32 84.22 82.42 81.72 81.52 81.42 7 79.46 76.96 78.26 78.86 79.96 81.26 81.86 80.76 78.96 78.26 78.06 77.96 8 78.9 76.4 77.7 78.3 79.4 80.7 81.3 80.2 78.4 77.6 77.5 77.2 9 74.22 72.62 73.02 73.62 74.72 76.02 76.62 75.52 73.72 72.92 72.82 72.72 10 78.69 76.69 76.49 77.09 78.19 79.49 80.09 78.99 77.19 76.39 76.29 76.19 11 77.71 75.91 75.51 76.11 77.21 78.51 79.11 78.01 76.21 75.41 75.31 75.21 12 80.69 78.59 78.49 79.09 80.19 81.49 82.09 80.99 79.19 78.49 78.29 78.19 13 81.07 78.97 78.87 79.47 80.57 81.87 82.47 81.37 79.57 78.87 78.67 78.57 14 76.1 74 73.9 74.5 75.6 76.9 77.5 76.4 74.6 73.9 73.7 73.6 15 75.25 73.15 73.05 73.65 74.75 76.05 76.65 75.55 73.75 73.05 72.85 72.55 16 75.74 73.64 73.54 74.14 75.24 76.54 77.14 76.04 74.24 73.44 73.34 73.24 17 70.89 69.89 68.19 67.79 70.19 76.39 77.79 76.69 72.89 71.09 70.69 70.39 18 75.19 74.09 73.99 74.19 74.89 76.39 77.19 76.59 74.79 74.09 73.79 73.49 19 72.58 71.08 70.38 70.98 73.38 74.98 76.08 74.78 72.18 71.38 70.88 70.68 20 68.86 66.76 66.36 67.46 69.66 72.76 74.26 73.16 70.36 68.66 67.96 67.16 21 68.71 67.71 66.51 67.11 69.61 73.91 75.81 74.71 70.91 69.21 68.11 67.71 22 73.03 71.73 71.43 71.83 74.03 75.23 75.93 74.83 73.03 72.33 71.73 71.53 23 80.86 78.96 78.66 79.26 81.66 84.56 85.46 84.36 82.56 80.86 80.26 80.06 24 73 70.9 70.8 71.4 73.9 76.5 77.7 76.6 74.8 73 72.5 72.3 25 74.44 72.34 72.24 72.84 75.34 77.74 79.14 78.11 76.24 74.84 74.54 74.34 26 70.37 68.27 68.17 68.77 71.27 73.57 75.07 73.97 72.17 70.97 70.17 69.97 27 72.7 71.1 70.5 71.1 73.6 76.1 77.4 76.3 74.5 73.9 71.9 71.7 28 71.36 69.86 69.16 69.76 72.16 74.46 75.96 74.86 73.06 72.06 70.66 70.36 29 76.94 74.84 74.74 75.34 77.74 79.64 81.54 80.44 78.94 77.64 76.54 76.24 30 74.48 72.38 72.28 72.88 75.28 77.18 79.08 77.98 75.76 75.18 74.58 74.38 31 75.48 73.38 73.28 73.88 76.28 78.18 80.08 78.98 76.3 75.68 75.18 74.88 32 69.78 67.68 67.58 68.18 70.58 72.48 74.38 73.28 71.48 70.48 69.68 69.38 American Academic Scientific Research Journal for Engineering, Technology, and Sciences (ASRJETS) - Volume 103, No 1, pp 268-289 289 33 80.7 78.6 78.5 79.1 81.5 83.9 85.3 84.2 82.4 81.37 80.5 80.3 34 74 73.2 72.5 73.1 74.5 77.4 78.3 77.2 75.4 73.55 73.2 73 35 83.33 81.93 81.13 82.23 84.13 86.03 87.93 86.83 84.03 82.33 82.03 81.83 36 76.54 74.44 74.34 74.94 77.44 79.24 81.24 80.14 77.34 75.64 75.44 75.24 37 77.18 75.08 74.98 75.58 78.08 79.88 81.88 80.78 77.08 76.58 75.88 75.68 38 81.64 79.54 79.44 80.04 82.54 85.04 86.34 85.24 83.44 82.34 81.54 81.24 39 83.3 82 81.1 81.7 84.2 86 88 86.9 85.1 84 83.6 83.3 40 77.61 76.21 75.41 76.01 78.41 80.31 82.21 81.11 79.31 77.61 77.11 76.99 41 78.41 76.31 76.21 76.81 79.21 81.61 83.01 81.91 79.11 77.91 77.31 77.01 42 77.07 74.97 74.87 75.47 77.87 79.77 81.67 80.57 78.77 76.07 75.67 77.77