Image-based air quality classification: superior performance of a VGG16 feature extraction and random forest pipeline

Abstract

Air pollution is a critical threat to environmental sustainability and public health, particularly in densely populated countries such as India and Nepal. This paper presents a hybrid pipeline that combines deep learning feature extraction with traditional machine learning classification for image-based air quality assessment. The method employs a two-stage pipeline: feature extraction using a pre-trained VGG16 CNN and classification using SVM and Random Forest (RF) algorithms. The model breaks down air quality into six levels: Good, Moderate, Unhealthy for Sensitive Groups, Unhealthy, Very Unhealthy, and Severe. A publicly available dataset of photographs taken at multiple urban sites in India and Nepal was used to conduct the test. Experimental results show that the Random Forest classifier achieves 97 percent accuracy, compared to 95 percent for the SVM classifier, on the test data. The effectiveness of the method is confirmed by a large-scale evaluation based on precision, recall, F1-score, and ROC-AUC metrics. The findings indicate a significant potential for combining deep feature extraction with traditional machine learning algorithms to create scalable environmental monitoring systems for real-world applications.