TRAQID - Traffic-Related Air Quality Image Dataset

PM in air affects optical images primarily through light scattering, including Rayleigh and Mie scattering [20]. This interaction is described by the Beer-Lambert law:

In this equation, I represents the observed hazy image, t is the transmission from scene to camera, J denotes scene radiance, and A is the airlight color vector. The first term, t(x, y)J(x, y), represents direct transmission of scene radiance to the camera, while the second term, (1 − t(x, y))A, describes airlight - ambient light scattered by air molecules and PM into the camera. This model assumes constant atmospheric and lighting conditions, although these factors may vary with weather, solar position, time of day, and season. Moreover, both J and A are influenced not only by meteorological conditions and solar position but also by PM distribution and concentration.

Color information plays a crucial role in PM estimation based on light scattering principles. Rayleigh scattering, predominant when particles are significantly smaller than the wavelength of light, exhibits strong wavelength dependence (λ^{− 4}), contributing to the sky's blue appearance. Conversely, Mie scattering occurs when particle sizes are comparable to light wavelengths, producing a white glare around the sun in particulate-laden air. The combined effects of Rayleigh and Mie scattering modulate the brightness and color saturation of outdoor images. Consequently, color and brightness information encapsulate particle concentration and size data, serving as distinctive features for PM estimation. This relationship between optical characteristics and atmospheric particulate content forms the basis for image-based air quality assessment methodologies.

Limited efforts have been devoted to predicting AQI using image-based methods, particularly in the aftermath of the COVID-19 pandemic, with a surge in research activity in this field. Prominent companies such as Google [27], Microsoft [32], IBM [24], and others have been actively engaged in this domain, employing advanced machine learning techniques to monitor air pollution. Researchers globally utilize images captured from stationary or mobile sources to analyze PM concentrations (PM_2.5 and PM₁₀) and predict AQI [15, 17, 19, 22, 25, 38]. However, most of the research in this field relies on custom datasets tailored to specific studies.

Dataset: Liu et al.[19] gathered 6587 images from three cities, Beijing and Shanghai in China and Phoenix in the U.S., from fixed scenes. The corresponding PM_2.5 data was obtained from local U.S. consulates, while weather and geographical data for the three cities were sourced from online websites. Kow et al.[17] gathered 3549 images taken in Kaohsiung, Taiwan, accompanied by corresponding air quality data, including PM_2.5, PM₁₀, and AQI, retrieved from nearby air quality monitoring stations. The dataset includes images captured during both daytime and nighttime. Zhang et al.[38] collected a total of 1241 sky images from a stationary location in Lanzhou, China, spanning from 2018 to 2019, and collected corresponding PM_2.5, PM₁₀ and gaseous pollutants data from the nearby monitoring station. In recent studies, Nilesh et al.[25] curated a dataset comprising 5048 images captured from roads of Hyderabad, India using a vehicular setup. Each image was paired with corresponding weather data and PM values obtained from sensors. The data collection occurred over two distinct seasons within a single year. Mondal et al.[22] used images captured from smartphone cameras to predict PM_2.5 concentration. They collected a custom dataset comprising 1818 images captured within Dhaka, Bangladesh, spanning from 2020 to 2022. The PM_2.5 concentration data corresponding to the images was obtained from the local US consulate, which releases hourly updates of the AQI. However, the dataset lacks features such as temperature and humidity, which play an effective role in determining the air quality.

AQI Estimation: Using either custom-collected or publicly available datasets, many researchers have used traditional and advanced deep learning methods to estimate the AQI and PM values. Liu et al.[19] utilized light attenuation and color information as significant image features for estimating PM levels from images. Additionally, machine learning models were employed to predict PM_2.5 values. Mondal et al.[22] utilized a custom-made CNN model with a total of 4.8 million parameters, featuring a single neuron in the final fully connected layer, to predict the PM_2.5 value. Kalajdjieski et al.[15] employed a multimodal approach, combining InceptionV3 [35] for image features and an MLP for weather data features. They extended the Inception architecture with a new sub-model path processing weather data through a three-layered MLP. The MLP output is concatenated with the Inception model's output in the final fully connected layer, feeding into a softmax layer to predict the AQI category. Zhang et al.[38] presented AQC-Net, a deep convolutional neural network model based on ResNet [13], aimed at classifying AQI categories. They integrated a self-supervision module called the Spatial and Context Attention block (SCA) into the ResNet18 architecture. This addition enabled encoding of broader scenes into local features and aggregation of spatial context information to enhance specific scene details within each channel. Nilesh et al.[25] employed a combination of deep learning and machine learning methods to predict AQI categories. They utilized YOLOv5 [8] to identify vehicles contributing to pollution in input images. A visibility metric was computed using BRISQUE [21] to assess image clarity. Image features (vehicle count and visibility score) were concatenated with weather features (temperature and humidity) to create feature vectors for each sample. Subsequently, a Random Forest (RF) [4] was trained to classify images into various AQI categories.

We introduce TRAQID - Traffic-Related Air Quality Image Dataset, designed for predicting AQI through image analysis. Our key contributions are:

1) The TRAQID dataset, comprising 26,678 samples from 70 hours of video capture, is the first of its kind to feature both front and rear traffic scene images alongside co-located sensor data. It uniquely combines vehicular traffic observations with essential environmental parameters such as temperature, relative humidity, and PM concentrations. This comprehensive approach captures the variability in urban environments, including diverse traffic patterns, building structures, and geographical features, providing a rich resource for analyzing the interplay between traffic dynamics and air quality.

2) The dataset incorporates AQI measurements categorized into six distinct ranges from “Good” to “Severe”, aligning with the classification schema outlined by the CPCB, India. It comprises a diverse collection of images containing both day and night settings across three primary seasons - Summer, Monsoon and Winter, spanning three distinct years.

Table 1 compares TRAQID with previously collected datasets, highlighting its unique features and advantages. Figure 1 showcases sample images, demonstrating the dataset's diversity across AQI categories, views, and environmental conditions. This comprehensive approach ensures that TRAQID captures various environmental conditions and traffic scenarios, making it a valuable resource for developing robust air quality estimation models.

TRAQID dataset creation summary

The data collection vehicle is equipped with two dashboard cameras as shown in Figure 2(a): 1. DDPAI Mola N3 [6] employed for front image capture, features a 5 MP CMOS sensor capable of recording in 2688 × 1944 ultra HD resolution, and 2. DDPAI X2 Pro [7] utilized for rear image capture, offers a 120° lens at the rear, ensuring a broad field of view while recording at 1920 × 1080 resolution. Both cameras record at a frame rate of 30 fps. These features facilitate high-quality image acquisition, enhancing data accuracy in urban settings.

The AQ Node is a custom device that measures PM concentrations and weather parameters at a frequency of 5 seconds. It includes an EspressIF ESP32 microcontroller unit (MCU) [33], a Nova SDS011 PM sensor [34] to measure PM_2.5 and PM₁₀ concentrations, and a BME280 sensor [11] to measure temperature and relative humidity. Additionally, an Aeroqual S500 [18], a reference device with a primary function of calibrating the PM sensors, was deployed, which captures a data point at a 1-minute frequency. The Nova PM sensor can measure particle size from 0.3 to 10 μm with a measuring range of 0.0 to 999.9 μg/m³. The AQ Node operates at a temperature range from -40°C to +125°C and a humidity range from 0% to 100% [28]. This equipment was mounted on the data collection vehicle to collect air quality data, capturing not only PM values but also the temperature and relative humidity, which are essential for assessing overall air quality [30].

The data collection campaign was conducted in the twin cities of Hyderabad and Secunderabad from October 2022 to July 2024, using the data collection vehicle shown in Figure 2(a). This vehicle travelled around the urban agglomeration, covering almost 2000 km at an average speed of 35 km/hr. The dataset collected in this campaign includes images extracted from 70+ hours of video, recorded on 20 distinct days across 6 different months in 3 years, ensuring diverse representation of environmental and traffic conditions in the twin cities. The specific route taken is detailed in Figure 2(b), with the paths plotted on OpenStreetMap [26] using GPS coordinates obtained from a mobile phone during the collection. The primary focus remained on collecting data within the city, with careful route planning to keep the vehicle within city limits. Samples were systematically collected throughout three distinct seasons: 1. Monsoon (October 2022, July 2024) 2. Winter (January, February and December 2023) and 3. Summer (March 2024) showcasing diverse climatic conditions.

Significance of City-Wide Coverage: The TRAQID dataset represents a significant advancement in air quality monitoring by providing comprehensive coverage of the twin cities of Hyderabad and Secunderabad. Capturing data across an entire urban agglomeration is a formidable task with several advantages. Firstly, it ensures the representation of diverse micro-environments within the cities, including residential areas, commercial zones, industrial sectors, and varying traffic densities. This diversity is crucial for developing robust air quality estimation models that can generalize across different urban settings. Secondly, the city-wide approach allows for capturing spatial variations in air quality, which can be significant even within short distances in urban areas due to localized pollution sources and varying topography. Moreover, the comprehensive nature of TRAQID facilitates the study of domain adaptation techniques for intercity air quality modelling. By capturing the full spectrum of urban environments, TRAQID provides a rich foundation for developing models that can potentially adapt to other urban areas, paving the way for more generalized air quality estimation systems across diverse city landscapes.

Image Data Preprocessing: The cameras mounted on the vehicle recorded 1-minute videos at a resolution of 1920 × 1080 and a frame rate of 30 fps. To align with the sampling rate of the sensors, we sampled images from the videos every 5 seconds to eliminate any repetitive frames and ensure consistency between the images and corresponding sensor data. Further, we filtered out outlier images manually, such as those affected by high amounts of headlight glare and extremely low illumination. The images were resized to 640 × 360 dimensions, making them compatible with deep learning architectures. To maintain dataset anonymity, we blurred both the number plates and faces within the images. Green-coloured number plates (indicating electric vehicles) were blurred while preserving their colour.

Sensor Data Preprocessing: To ensure the reliability of our air quality measurements, we implemented a rigorous calibration process for our low-cost PM sensors, following the method described by Ayu et al.[28]. This process involved comparing sensor readings against a reference device (Aeroqual S500) in a controlled environment, applying linear regression for calibration, and using the interquartile range (IQR) method to eliminate outliers. The resulting calibrated sensor data was then synchronized with our image captures, creating what we term “co-located” data points. This co-location approach is crucial for our vision-based air quality estimation task, as it establishes a direct temporal and spatial correspondence between the visual characteristics of traffic scenes and their associated air quality measurements. By ensuring this tight coupling between visual data and air quality parameters, we provide a solid foundation for developing and evaluating computer vision algorithms that can infer air quality from image content.

Label Histogram

Dataset distribution wrt AQI

The proposed dataset contains a comprehensive compilation of 26,678 data samples, precisely gathered to ensure diversity across various environmental contexts. The dataset contains 13789 images obtained during the daytime (6 AM - 6 PM) and 12889 images obtained at nighttime (6 PM - 6 AM), which enables the possibility of predicting the AQI at different times of the day. This is especially useful for studying pollutants primarily observed at a particular time, such as smog, usually observed at night. The AQI values of the data samples are computed using the pollutants PM_2.5 and PM₁₀ concentration values as defined by the CPCB, India. Further, the AQI levels are categorized into six distinct classes which are as follows: 1. Good (0 - 50) 2. Satisfactory (51 - 100) 3. Moderate (101 - 200) 4. Poor (201 - 300) 5. Very Poor (301 - 400), and 6. Severe (> 400). As shown in Figure 2(d), each data sample comprises: 1. Front and rear images 2. Temperature (°C) 3. Relative humidity (%) 4. Season 5. Timestamp 6. PM_2.5 concentration value 7. PM₁₀ concentration value 8. AQI Value 9. AQI category (according to the CPCB standards). The front and rear images, with 640 × 360 × 3 dimensions, help understand the traffic scenario's vehicular dynamics. Temperature and relative humidity are scalar, and the season is categorical (summer, winter, monsoon). The timestamp contains the date and time of the sample captured. Another notable feature of the dataset is the sequential organization of the samples. Since the data was collected on 20 different dates, the data samples collected each day form a sequence, representing the progression of air quality as the data collection vehicle traverses the streets. For detailed dataset directory structure, refer to the supplementary material.

Figure 3 presents the histogram of labels for both day and night data samples. Figure 4 displays the dataset distribution across AQI categories and seasons, highlighting a deficiency in the “good” category compared to other categories. This suggests that AQI is generally adverse in traffic scenarios, with most data samples ranging from “satisfactory” to “moderate”, reflecting the typical air quality in Hyderabad city. The “poor”, “very poor”, and “severe” categories are significant, pointing to local events like construction activities, industrial operations, and open burning as contributing factors.

Analysis of air quality data reveals distinct seasonal and diurnal variations in pollutant concentrations and AQI categories. As shown in Table 2, monsoon season exhibited the best air quality with 643 ‘Good’ and 7,296 ‘Satisfactory’ AQI days, and the lowest mean AQI of 109. Conversely, summer presented the worst conditions, with zero ‘Good’ AQI days, the highest mean PM₁₀ (209 μg/m³), and highest mean AQI (190). Winter recorded 601 ‘Severe’ AQI days and the highest mean PM_2.5 (74 μg/m³). Diurnal patterns showed poorer daytime air quality (mean AQI 164) compared to nighttime (mean AQI 121), with higher frequencies of ‘Poor’, ‘Very Poor’, and ‘Severe’ categories during the day. PM₁₀ levels consistently exceeded PM_2.5 across all temporal divisions. Summer recorded the highest mean temperature (36° C) and lowest humidity (23%), while monsoon showed moderate temperatures (29° C) and high humidity (49%). These findings highlight the complex relationship between meteorological factors and air quality.

As mentioned above, the following experiments were conducted to benchmark the specified methodologies on the TRAQID dataset. As defined in Mondal et al.[22], a modified version of the standard CNN framework was used to estimate PM_2.5, PM₁₀, AQI value and AQI categories. This architecture consisted of a total of 19 layers, encompassing both convolutional and fully connected layers, with approximately 4.8 million parameters in total. Rectified Linear Unit (ReLU) activation function was used, and the model was trained for 100 epochs with early stopping criteria and a learning rate of 1e^{− 3}.

According to Zhang et al.[38], the AQC-Net was employed for both regression and classification tasks. The architecture contains a total of 0.03 million parameters. Training was conducted for 100 epochs with an initial learning rate of 1e^{− 2}, subsequently reduced by a factor of 0.1 every 30 epochs.

As mentioned in Nilesh et al.[25], YOLOv5 was used to obtain vehicle counts from both rear and front images, which were subsequently aggregated to form the image features, while the weather features remained unchanged. However, since our dataset includes night images, the BRISQUE methodology was ineffective in determining image visibility scores and was excluded. For regression, RF regressor models predicted PM concentrations and AQI values, while RF classification models predicted AQI categories.

As defined in Kalajdjieski et al.[15], a multimodal architecture was implemented that leverages InceptionV3 [35] to extract image features, while a three-layer MLP with 64 neurons in each layer is utilized to extract weather features (including temperature, relative humidity, season, and timestamp categorized as day/night). The output feature vector from the MLP is concatenated with the output from InceptionV3 at the fully connected layer, which is then fed into a softmax layer for AQI category prediction. For regression tasks, instead of a softmax layer, a single neuron output layer is trained to predict PM or AQI values. The model contains approximately 24 million parameters. The model was trained for 100 epochs with a learning rate of 1e^{− 3}, using the early stopping criteria.

For all the methodologies implemented above, the Mean Squared Error (MSE) loss function was used for regression tasks, while the cross-entropy loss function was utilized for classification tasks. In [15, 22, 38], the TRAQID dataset's front and rear images were sequentially processed by their respective CNN model. Subsequently, the outputs were concatenated and passed to the fully connected layer. PyTorch [29] was utilized for all implementations.

The results of the methodologies on the TRAQID dataset are presented in Table 3. It is clearly observed that the conventional CNN models used by [15, 22] fail to grasp the intricacies of the task at hand. This is primarily due to the fact that standard CNN models may struggle to incorporate the diverse factors affecting air quality, including industrial activities, construction, environmental variations, and geographical features, into their predictions. AQC-Net [38] stands out among the methods, demonstrating superior performance in three tasks which is PM₁₀ estimation, AQI value estimation and AQI category estimation. This highlights the effectiveness of attention-based methods in identifying pollution-contributing objects in images. However, the top F1-score achieved for the AQI category estimation task is just 0.74, highlighting the significant impact of dataset size and diversity on the complexity of the task. Inclusion of nighttime images brings challenges like low illumination and glare through headlights which makes the task at hand more complicated in compare to the datasets which only contains daytime images.

TRAQID Challenges

Nilesh et al.[25] achieved the highest score in PM_2.5 estimation task, demonstrating that incorporating vehicle counts alongside weather parameters can aid in air quality estimation. This can be viewed as a specific instance of attention mechanism, where vehicle emissions constitute a significant pollution source on roads and streets. It suggests a positive correlation between traffic volume and pollution levels. However, the model achieved only a 0.71 F1-score for the AQI category estimation task, representing a 10% decrease for the same task compared to their proposed dataset. This reduction may be attributed to the improved season diversity in our dataset, feature not present in previous datasets. Additionally, the presence of unstructured traffic further complicates the task. Figure 5 shows some of the classification challenges associated with the dataset. The top and bottom left figures show data samples having high ground truth AQI value. However, the best-performing model for this task predicts a very low AQI value, indicating a failure to detect pollution caused by nearby bridge construction. Similarly, the model struggles with low illumination in the top right image and unstructured traffic scenario in the bottom right image. Please refer to the supplementary material for further insights into the dataset challenges.

To investigate seasonal and diurnal variations, we segmented the dataset into multiple splits. For seasonal analysis, we created three subsets: 1. Monsoon, 2. Winter, and 3. Summer. For diurnal analysis, we divided the data into two categories: 1. Day and 2. Night. Each of these split datasets was further partitioned into training (80%) and validation (20%) sets. We then trained the best-performing model for all four defined tasks to each of these datasets. Table 4 presents the results for each dataset.

The challenge of modeling summer data is evident, even for the best-performing model, primarily due to the fluctuation in temperature and humidity during summers. Conversely, the model achieved its best performance on monsoon data, likely because a majority of the data samples fall within the 0 - 200 AQI range, attributed to the rainfall during the monsoon season. As for the day/night comparison, the model exhibited superior performance in predicting nighttime air quality parameters compared to daytime. This can be attributed to several factors: reduced human activity and more stable meteorological conditions at night, resulting in lower variability in air quality (as shown in Fig. 3 and Table 2); a more stable atmospheric boundary layer, leading to consistent pollutant concentrations; and fewer extreme AQI categories (“Poor”, “Very Poor”, “Severe”) at night, which aligns with the model's tendency to perform better on less extreme values. The reduced influence of rapid fluctuations from daytime events (e.g., rush hour traffic, sudden weather changes) likely enhances the model's nighttime predictions. Analysis of camera perspectives revealed inferior performance when using single views, with front-view models performing marginally better than rear-view models. This indicates the importance of utilizing both camera perspectives for optimal air quality prediction.

TRAQID dataset grad cam

To visualize and interpret the AQC-Net [38] decision-making process for AQI category estimation, we employed Gradient-weighted Class Activation Mapping (GradCAM) [31]. The GradCAM technique was applied to the final convolutional layer of the trained model, generating heatmaps that highlight the regions of input images most influential in determining the predicted AQI category. These visualizations were generated for multiple images across different AQI categories to analyze how the model's focus shifts with varying air quality conditions. As illustrated in Fig. 6, the model consistently focuses on vehicles, particularly their exhaust areas, across all AQI categories, indicating its recognition of vehicular emissions as a crucial air quality factor. For “Good”, “Satisfactory”, and “Moderate” categories, the model attends to urban structures such as flyovers, buildings, and bridges. However, as AQI deteriorates from “Poor” to “Severe”, attention shifts towards open areas and light sources, potentially seeking indicators of haze or atmospheric clarity. Notably, in “Severe” AQI images, the model considers broader image portions, suggesting an analysis of overall atmospheric conditions rather than specific pollution sources. These observations suggest that the model has learned to associate air quality with complex set of visual cues related to traffic patterns, urban infrastructure, and atmospheric conditions.