Modelling of Urban Near-Road Atmospheric PM Concentrations Using an Artificial Neural Network Approach with Acoustic Data Input

Mobile monitoring and spatial prediction of black carbon in Cairo, Egypt

This study harnesses the power of mobile data in developing a spatial model for predicting black carbon (BC) concentrations within one of the most heavily populated regions in the Middle East and North Africa MENA region, Greater Cairo Region (GCR) in Egypt. A mobile data collection campaign was conducted in GCR to collect BC measurements along specific travel routes. In total, 3,300 km were travelled across a widespread 525 km of routes. Reported average BC values were around 20 µg/m³, announcing an alarming order of magnitude value when compared to the maximum reported values in similar studies. A bi-directional stepwise land use regression (LUR) model was developed to select the best combination of explanatory variables and generate an exposure surface for BC, in addition to a number of machine learning models (random forest gradient boost, light gradient boost model (LightGBM), Keras neural network (NN)). Data from 7 air quality (AQ) stations were compared-in terms of mean square error (MSE) and mean absolute error (MAE)-with predictions from the LUR and the NN model. The NN model estimated higher BC concentrations in the downtown areas, while lower concentrations are estimated for the peripheral area at the east side of the city. Such results shed light on the credibility of the LUR models in generating a general spatial trend of BC concentrations while the superiority of NN in BC accuracy estimation (0.023 vs 0.241 in terms of MSE and 0.12 vs 0.389 in terms of MAE; of NN vs LUR respectively).

Roadside vehicle particulate matter concentration estimation using artificial neural network model in Addis Ababa, Ethiopia

Currently, vehicle-related particulate matter is the main determinant air pollution in the urban environment. This study was designed to investigate the level of fine (PM_2.5) and coarse particle (PM₁₀) concentration of roadside vehicles in Addis Ababa, the capital city of Ethiopia using artificial neural network model. To train, test and validate the model, the traffic volume, weather data and particulate matter concentrations were collected from 15 different sites in the city. The experimental results showed that the city average 24-hr PM_2.5 concentration is 13%-144% and 58%-241% higher than air quality index (AQI) and world health organization (WHO) standards, respectively. The PM₁₀ results also exceeded the AQI (54%-65%) and WHO (8%-395%) standards. The model runs using the Levenberg-Marquardt (Trainlm) and the Scaled Conjugate Gradient (Trainscg) and comparison were performed, to identify the minimum fractional error between the observed and the predicted value. The two models were determined using the correlation coefficient and other statistical parameters. The Trainscg model, the average concentration of PM_2.5 and PM₁₀ exhaust emission correlation coefficient were predicted to be (R² = 0.775) and (R² = 0.92), respectively. The Trainlm model has also well predicted the exhaust emission of PM_2.5 (R² = 0.943) and PM₁₀ (R² = 0.959). The overall results showed that a better correlation coefficient obtained in the Trainlm model, could be considered as optional methods to predict transport-related particulate matter concentration emission using traffic volume and weather data for Ethiopia cities and other countries that have similar geographical and development settings.

Application of Machine Learning for the in-Field Correction of a PM2.5 Low-Cost Sensor Network

Many low-cost sensors (LCSs) are distributed for air monitoring without any rigorous calibrations. This work applies machine learning with PM_2.5 from Taiwan monitoring stations to conduct in-field corrections on a network of 39 PM_2.5 LCSs from July 2017 to December 2018. Three candidate models were evaluated: Multiple linear regression (MLR), support vector regression (SVR), and random forest regression (RFR). The model-corrected PM_2.5 levels were compared with those of GRIMM-calibrated PM_2.5. RFR was superior to MLR and SVR in its correction accuracy and computing efficiency. Compared to SVR, the root mean square errors (RMSEs) of RFR were 35% and 85% lower for the training and validation sets, respectively, and the computational speed was 35 times faster. An RFR with 300 decision trees was chosen as the optimal setting considering both the correction performance and the modeling time. An RFR with a nighttime pattern was established as the optimal correction model, and the RMSEs were 5.9 ± 2.0 μg/m³, reduced from 18.4 ± 6.5 μg/m³ before correction. This is the first work to correct LCSs at locations without monitoring stations, validated using laboratory-calibrated data. Similar models could be established in other countries to greatly enhance the usefulness of their PM_2.5 sensor networks.

Modelling of Urban Air Pollutant Concentrations with Artificial Neural Networks Using Novel Input Variables

Since operating urban air quality stations is not only time consuming but also costly, and because air pollutants can cause serious health problems, this paper presents the hourly prediction of ten air pollutant concentrations (CO2, NH3, NO, NO2, NOx, O3, PM1, PM2.5, PM10 and PN10) in a street canyon in Münster using an artificial neural network (ANN) approach. Special attention was paid to comparing three predictor options representing the traffic volume: we included acoustic sound measurements (sound), the total number of vehicles (traffic), and the hour of the day and the day of the week (time) as input variables and then compared their prediction powers. The models were trained, validated and tested to evaluate their performance. Results showed that the predictions of the gaseous air pollutants NO, NO2, NOx, and O3 reveal very good agreement with observations, whereas predictions for particle concentrations and NH3 were less successful, indicating that these models can be improved. All three input variable options (sound, traffic and time) proved to be suitable and showed distinct strengths for modelling various air pollutant concentrations.

Application of Artificial Neural Networks for Noise Barrier Optimization

In the modern world, noise pollution continues to be a major problem that impairs people’s health, and road traffic is a primary contributor to noise emissions. This article describes an environmental impact study of the noise generated by the reconstruction of an urban section of a highway. Noise maps were calculated, and an environmental impact matrix was generated to determine the environmental impact of this reconstruction. The implementation of noise barriers was simulated based on these noise maps, and the effectiveness of the barriers was evaluated using Artificial Neural Networks (ANNs) combined with Design of Experiments (DoE). A functional variable significance analysis was then made for two parameters, namely, the coefficient of absorption of the barrier material and the barrier height. The aim was to determine the influence of these parameters on sound attenuation and on the formation of acoustic shadows. The results obtained from the ANNs and DoE were consistent in demonstrating that the absorption coefficient strongly influences the noise attenuation provided by noise barriers, while barrier height is correlated with the formation of larger areas of acoustic shadow. The environmental impact matrix also indicates that the existence of noise pollution has a negative effect on the environment, but that this impact can be reversed or minimized. The application of simulated noise barriers demonstrated that noise levels can be reduced to legally acceptable levels.