An ensemble deep learning approach for air quality estimation in Delhi, India

South Asian megacities, notably Delhi, are significant contributors to air pollution, driven by factors such as population density, vehicular emissions, and industry, posing serious health and environmental challenges. This study addresses the accurate air quality estimation problem in Delhi, proposing an ensemble model named En3C-AQI-Net. Leveraging a fine-tuned Data Efficient Image Transformer (DeiT) with outdoor images, a Convolutional Neural Network (CNN) incorporating dark-channel prior, and a 1-dimensional CNN trained with meteorological features, the model combines predictions using weighted average ensemble learning to classify images into six Air Quality Index (AQI) classes and estimate AQI values. Ensemble learning enhances model performance by combining diverse predictions, increasing stability, and overall accuracy. To train and validate the proposed model, an image dataset labeled with ground-truth AQI values is collected from Delhi, India, termed ‘AirSetDelhi’. It consists of 21,620 single-scene daytime and nighttime images, unevenly distributed in six AQI categories. Experimental results show that En3C-AQI-Net outperforms pre-trained CNN models, achieving 89.28% accuracy and a Cohen Kappa score of 0.856 for AQI classification, with an RMSE of 47.36 and an R² value of 0.861 for AQI estimation, demonstrating efficacy in both tasks. Accuracy measures overall correctness, Cohen Kappa assesses model agreement, RMSE quantifies average prediction error, and R² value gauges the proportion of variance explained in a model. The En3C-AQI-Net architecture captures diverse and complementary features, demonstrating efficiency as an alternative feasible approach for air quality estimation in the challenging context of Delhi. However, limitations include potential applicability constraints and real-time deployment challenges.