Deep Learning Based Air Visibility Level Detection

With increased urbanisation and man-made activities,the problem of air pollution is becoming more and more prominent,with hazy weather being a frequent occurrence.The reduction in visibility caused by haze not only has many negative effects on people’s lives,but also on transport,environmental monitoring and weather forecasting.Traditional methods of air visibility detection are constrained by hardware conditions and weather conditions,resulting in limited detection effectiveness.Therefore,deep learning-based detection of air visibility levels has become a current research need.In this paper,we use deep learning algorithms to train and learn from a large amount of data to effectively improve the accuracy and stability of air visibility level detection,and the main research content is as follows:First,the air visibility class detection based on convolutional neural network model is studied.The convolutional neural network is used to learn the characteristics of haze images,and the true value is calculated based on the images using a large amount of human effort by calculating the mean value as the basis for air visibility class classification.By training and validating the data set,and using the optimal network model to detect the visibility level of the test set data,the accuracy of 84.5% is obtained,which demonstrates the effectiveness of using convolutional neural network to detect the air visibility level.Second,to reduce the training time and speed up the fitting of the network model to the data.A migration learning based VGG neural network for air visibility level detection is investigated.The model is retrained and tested by loading the trained VGG16 model weights on the Image Net dataset.The accuracy of the migration learning-based VGG neural network for air visibility detection is 86.5%,which shows that this adaptation is feasible.Finally,for the characteristics of haze images,the underlying information of haze images can be of great help in the detection of visibility.In contrast,the traditional VGG network,due to its deep network structure,may lose the underlying detailed information after repeated feature extraction.Therefore,the final study of this paper designs an improved neural network model Optimization-VGG based on the traditional VGG neural network model and with reference to the concept of residual structure,combining the low-level feature map and the high-level feature map as the feature information of the haze image,mainly by globally averaging the detailed features of the bottom module and then stacking them with the original high-level feature information.The extraction of the underlying detail information is enhanced to enable better feature extraction and to improve the accuracy of discriminating between different air visibility levels.Finally,through the analysis of the experimental results,the improved network model achieves an accuracy of 90.5% in the detection of visibility levels of haze images,which meets the practical requirements.Compared with the traditional network model,the Optimization-VGG neural network model has significantly improved the detection of air visibility levels.