Aerosol can not only affect urban air quality, but also reduce atmospheric visibility, hinder people’s visual range, affect urban traffic, and bring great hidden dangers to traffic safety. Aerosol also plays an important role as a medium in the transmission of infectious diseases, which can cause or aggravate respiratory diseases, and seriously affect people’s health. At present, the commonly used tools for aerosol observation include spaceborne lidar and ground-based lidar. Spaceborne lidar will be affected by weather, satellite trajectory, observation distance and other factors, while ground-based lidar can make long-term fixed-point observation with high detection accuracy. Due to its high spatial and temporal resolution and high precision, ground-based lidar can realize rapid response to local pollution sources and highly polluted air masses. It is widely used in the field of atmospheric environment monitoring to explore the spatiotemporal evolution of haze, dust, and planetary boundary layer (PBL). It can also trace pollution to its source to provide technical support for environmental protection. However, the use of lidar horizontal scanning technology to trace the pollution source is mainly carried out manually. Thus, the relevant law enforcement departments cannot immediately investigate until the implementation of pollution hotspot identification. In this paper, deep learning image analysis technology is utilized in lidar horizontal scanning technology, and a technology that can automatically identify and plot pollution sources is developed based on the image classification and image segmentation algorithm of deep learning image analysis technology to achieve the fast identification of hotspots. The results show that different types of pollution sources can be automatically identified, such as point pollution, line pollution, and area pollution. The technology was tested in Dangshan County, Anhui Province. The results show that the pollution source identification accuracy can reach 91.5% and the pollution sources such as fireworks discharge source, dust, and incineration sources could be accurately identified.

Two algorithms, i.e., image classification and image segmentation based on Baidu PaddlePaddle platform, are used to identify pollution sources. Firstly, a training dataset from a large number of different scanning spectra of aerosol lidars is established. Then the lidar scanning spectra are classified by image classification method, which can be roughly divided into normal spectra, equipment fault spectra, meteorological anomaly spectra, noise anomaly spectra, etc. The normal spectra are selected for the next step to identify and extract pollution hotspots. Next, the image segmentation algorithm is used to outline the pollution hotspots and the pollution sources are classified according to the segmented pollution shapes, including point pollution, line pollution, and area pollution. Finally, based on the identified shape contour information and relevant geographic algorithms, the actual area of the contour is calculated.

By the image classification and image segmentation from lidar-derived training dataset, different types of pollution sources can be automatically identified, such as point pollution, line pollution, and area pollution. The effect diagram after image segmentation is shown in Fig. 4 and the different types of lidar scanning spectra identified by automatic identification technology are shown in Fig. 5. The technology which can automatically identify and plot the pollution source was tested in Dangshan County, Anhui Province during December 2022-January 2023. Based on the different characteristics of pollution sources, for example, the depolarization ratio of the fireworks discharge source between 0.07 and 0.10, that of the dust source between 0.1 and 0.15, and that of the incineration source between 0.05 and 0.08, the different pollution sources can be classified. Totally 106 pollution hotspots were automatically identified using this technology, of which 97 pollution sources were confirmed by manual audit, with an accuracy of 91.5%. Meanwhile, a total of 103 pollution sources were confirmed by manual audit and the probability that the technology could not identify the pollution source was 5.8%.

An automatic technology for pollution source identification based on the Baidu PaddlePaddle platform is developed. By the image classification and image segmentation from lidar-derived training dataset, different types of pollution sources can be automatically identified. Specifically, the technology which can automatically identify and plot the pollution source was tested in Dangshan County, Anhui Province. The results show that the accuracy can reach 91.5%. And the pollution sources such as fireworks discharge source, dust, and incineration source could be accurately identified.