Research On PM2.5 Fine-Grained Detection Algorithm Based On Deep Learning

Rapid industrialization lead to serious air pollution problems.PM2.5 particles,as a typical indicator of urban air quality monitoring,have aroused widespread concern.So far,the detection method by air pollution diffusion mechanism is complicated to model.The detection model based on remote sensing is unstable and lacks flexibility,which makes it impossible to complete the fine-grained detection of PM2.5.To solve these problems mentioned above,this thesis proposes a series of solutions to improve the accuracy and flexibility of PM2.5 fine-grained detection.The main contributions are summarized as follows:(1)A PM2.5 fine-grained detection framework based on mobile crowdsensing is designed.The framework involves three parts: data collection & preprocessing,visual model & meteorological model training,as well as the fusion of decision-level models.(2)A PM2.5 fine-grained detection solution based on outdoor sky images is proposed.Dark channel conversion was performed on the collected outdoor sky images to better reflect the level of haze.Through building residual neural network,a visual model based on outdoor sky images can be trained for PM2.5 fine-grained detection.(3)A fine-grained PM2.5 detection solution based on meteorological data is presented.Meteorological data & pollution data sets were collected through web crawlers.Through building GRU network,a meteorological model for fine-grained PM2.5 detection based on meteorological data can be obtained.(4)We make decision-level fusion of visual model and meteorological model to further enhance the accuracy of PM2.5 fine-grained detection,and thus improve the problems of unstable performance and poor robustness caused by single PM2.5 concentration detection model.This thesis trains the collected sky images and meteorological data to obtain the visual model & the meteorological model,then performs a fusion experiment on the two detection models.The experimental results show that: when the visual model is used alone,the regression error and the confusion matrix classification accuracy are 27.93 and 0.79,respectively;when the meteorological model is used alone,the regression error and the confusion matrix classification accuracy are 21.51 and 0.91;when the weight coefficient is 0.45,the regression error of the fusion model is reduced to a minimum of 7.86,and the confusion matrix classification accuracy is a maximum of 0.98.Compared with classical interpolation methods and other single detection models,the accuracy and flexibility of this proposed solution have been improved,and it can solve the problems of unstable performance and poor robustness of single detection model.