**1. Introduction**

The main component of haze is fine particulate matter (PM2.5), an organic compound of toxic substances such as heavy metals and carcinogens. It is the most harmful air pollution to the human body because it can directly enter the lungs [1–3]. In recent years, more and more people have paid close attention to its environmental damage and influence on the human body in large urban areas [4–6].

With the increased attention on haze, the harm it does to the human body is gradually being revealed. Therefore, the research on haze becomes deeper and more diverse [7–11]. Predominant questions involve the causes of haze, pollution composition, time distribution, regional distribution, and management programs. The research methods also span multiple disciplines such as chemical analysis, biological testing, economic development, and haze data mining.

Because of the improved human aerospace technology and remote sensing satellite technology in recent years, the cost of remote sensing satellite imagery has been reduced. It is more macroscopic than the traditional ground station monitoring data. The satellite images provide the information of the temporal and spatial changes of haze comprehensively and quickly [12–18]; therefore, researchers utilize remote sensing images in the monitoring and analysis of haze. Researchers often use remote sensing satellite images for the inversion of Aerosol Optical Depth (AOD) and further analyze meteorological features based on the correlation between aerosol depth and atmospheric pollutant concentrations. McGowan et al. [19] proposed a PM10 dust concentration of a 500 m vertical

**Citation:** Yin, L.; Wang, L.; Huang, W.; Tian, J.; Liu, S.; Yang, B.; Zheng, W. Haze Grading Using the Convolutional Neural Networks. *Atmosphere* **2022**, *13*, 522. https:// doi.org/10.3390/atmos13040522

Academic Editors: Duanyang Liu, Kai Qin, Honglei Wang and Rajasekhar Balasubramanian

Received: 18 January 2022 Accepted: 22 March 2022 Published: 25 March 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

profile measured during a regional dust event in western Queensland, Australia, based on MODIS Terra satellite data and a spatiotemporal analysis. Guo et al. [20] used a correlation analysis between the PM2.5 concentration ground haze monitoring stations in China during 2007 and 2008 and the AOD obtained from the satellite remote sensing image. They also discussed the feasibility of satellite remote sensing technology for estimating the haze concentration on the ground. Nordio et al. [21] used MODIS data to study the correlation between aerosol and PM10 concentration in Lombardy, Italy, and successfully used aerosol data to predict the haze concentration. Seo et al. [22] compared the aerosol depth based on ground monitoring stations and MODIS satellite images to PM10 concentrations in Seoul, Korea. It was concluded that MODIS images are more relevant than ground monitoring stations, especially in winter.

Another emerging approach to haze research is machine learning. Machine learning methods, especially the rapid development of neural networks, have shown researchers great potential to fit complex functions. Therefore, some studies have used haze detection data to train neural networks to predict haze concentrations. For example, Pérez et al. [23] used the neural network structure to predict and analyze the average concentration of haze in the San Diego area in the next few hours. Grivas et al. [24] optimized the structure and parameters of the neural network based on the previous Pérez study to predict the concentration of PM10. With the continuous optimization of the network, many scholars have used neural networks to predict and analyze the haze in time series [25–28].

Comparing the two haze analysis methods, estimating the concentration of haze pollutants, such as PM2.5 and PM10, using the inversion of AOD on remote sensing satellite images has a broader research background and relative higher precision. However, the data preprocessing for obtaining AOD from satellite images is very complicated and often requires steps such as radiation correction, geometric correction, and processing angle data for satellite images. By contrast, neural networks have powerful feature capture capabilities and flexible adjustment capabilities using these learned features. These characteristics allow neural networks to use simplified data processing compared with traditional inversion methods. Moreover, the neural network excels in complex function fitting, making it easy to fit uncertain function expressions or expressions with complex parameters. In the study of haze, PM2.5 concentration prediction is a complex function fitting problem, since the concentration of other major pollutants such as PM2.5 depends on many complicated meteorological and human factors.

This paper aims to simplify complex data processing in traditional AOD inversion methods by using the feature capture ability and complex function fitting ability of neural networks by training a haze-level classification network. We directly use remote sensing satellite images as input and use convolutional neural networks as training models to classify the level of haze concentration. This paper compares the results from two methods, one using a traditional AOD inversion method, and the other the proposed neural networks inversion method. Experiments show that the proposed network can reduce the manual inversion work, and also achieves good results in fitting the non-linear relationship between the data and the haze concentration level.
