Research On Air Quality Monitor Method Based On Transformer

Air quality monitor is an essential part of environmental protection.By collecting real-time air quality data in various regions and building a dynamic air quality testing network,regulatory authorities can objectively analyze the impact of air pollution on the environment and human health.Currently,the air quality monitor method used by air quality monitor stations in China requires complex physical and chemical sensors,which brings certain limitations in frequency and flexibility.With the development of deep learning technology in recent years,many air quality monitor models based on visual data have appeared.These models allow ordinary people with only cell phones to instantly obtain AQI(Air Quality Index)based on the images they capture.The participation of ordinary people will provide extensive image data with complex and similar content for air quality monitor tasks.This type of data classification is a fine-grained classification task,which fits the Transformer with the self-attention mechanism and pays more attention to local information.Therefore,we propose DOViT(Double Output Vision Transformer).The method will process the images taken and uploaded by the user using the mobile device in the cloud and then predict the local AQI.Compared with traditional methods,the present method dramatically improves flexibility and detection frequency.Due to the lack of public datasets in this research direction,this paper establishes the dataset GAOs-2(Get AQI in One shot-2),which contains 1,054 high-quality sky images captured by mobile devices for experiments and testing augmented in the process of use.The comparison experiments with popular air quality monitor methods prove that the DOViT proposed in this paper can accurately predict the AQI level.The DOViT dramatically increases the air quality monitor’s frequency and flexibility.It can become an auxiliary means for environmental protection departments to monitor pollution.Also due to the lack of public datasets,this paper explores the Transformer trained with self-supervised learning methods.It proposes the JiT(Jigsaw Vision Transformer)and GAOs-3(Get AQI in One shot-3)with 6,000 unlabeled high-quality sky images.The results show that the JiT achieves high accuracy using only a tiny labeled dataset,which means it has research value.