SmartAir: Enhancing Air Quality Classification with Deep Learning and Two-State Q-Learning

Abstract

Air pollution, especially PM 2.5, poses a growing global health threat, mainly from industrial activity, traffic, and wildfires. Traditional air quality monitoring systems are costly and lack broad coverage. This research proposes a low-cost alternative using a model that classifies air quality from outdoor images by combining Deep Learning with Reinforcement Learning. The model uses VGG19 (pre-trained on ImageNet) along with Local Binary Pattern (LBP) and RGB average values for feature extraction. A non-linear SVM classifier is enhanced with Q-Learning, which improves classification of difficult images through randomized actions: rotating images <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\pm 15$</tex> degrees or shifting them diagonally. Experimental results show that incorporating Q-Learning increased model accuracy from <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{96.11\%}$</tex> to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{97.29\%}$</tex>. This indicates that reinforcement learning helps the system adaptively correct misclassifications. The proposed model offers an efficient, accessible, and affordable tool for air quality monitoring, especially in areas lacking conventional AQI sensors.