TDPM-CNN: a comprehensive daytime and nighttime PM2.5 estimation method using multi-kernel convolutional neural networks
摘要
Air pollution has become a serious global issue, and the World Health Organization (WHO) considers it a significant threat to human health. In 2019, air pollution was responsible for the deaths of 4.2 million people, with PM2.5 being one of the major air pollutants. In recent years, the emergence of computer vision and deep learning has led to an increasing number of studies using convolutional neural networks (CNN) for estimating PM2.5 concentrations in images. However, many of these methods treat PM2.5 estimation as a classification problem, limiting their applicability to specific regions or periods. We propose a CNN model with dense connections and multi-kernel convolutional kernels for estimating the PM2.5 concentration in images. Since both daytime and nighttime images were used for PM2.5 concentration prediction, this model was named Throughout the Day PM2.5 Convolutional Neural Network (TDPM-CNN). The model takes as input the combination of the original image, high-frequency features capturing edge information such as building edges, and low-frequency features representing color distributions in the image. These three types of information are merged as a multi-channel feature, which is passed through seven multiple kernel convolutional blocks, where 3*3, 5*5, and 7*7 convolutional kernels are applied, followed by feature concatenation and pooling operations. Experiments were conducted using the TDPM-CNN model on both daytime and nighttime images. For daytime images, the Root Mean Square Error (RMSE) achieved