Image-Based Air Pollution Classification Using Deep Learning Techniques

Abstract

Air pollution is a considerable threat to human health and environmental sustainability. Traditional monitoring techniques involve sensor networks, which can be expensive, spatially constrained, and time-consuming to scale. This paper aims to examine a deep learning-based image classification technique to analyze air pollution levels using environmental imagery. A labelled dataset with varying levels of pollution intensity was used to train convolutional neural networks (CNNs) on visual indicators such as haze density, sky colour, and visibility. Several well-known architectures were assessed, namely: ResNet50, AlexNet, VGG16, VGG19, Xception, and InceptionV3. The findings indicate that the best model, which is invariant and accurate, was the model called Xception. To improve generalisation and robustness, regularisation techniques such as dropout, batch normalisation, and data augmentation were applied. Model performance was assessed using accuracy, F1-score, precision, and recall. The highest results were achieved by Xception, with 90.45% (test), 90.21% (train), and 88.31% (validation). VGG16 and VGG19 were the next highest performing models. Conversely, ResNet50 demonstrated the poorest performance across all metrics These findings highlight the potential of advanced CNN architectures as a cost-effective and scalable alternative to traditional sensor-based monitoring, providing valuable insights for smart city applications and sustainable urban planning