Research On Forecast Method Of Spatio-temporal Evolution Of Haze Based On CNN

In recent years,haze pollution has received widespread attention.The increasing number of ground-based monitoring stations and monitoring satellites has brought about more sporadic spatio-temporal data,at which time the traditional methods appear to be inefficient.With the development of machine learning and deep learning,researchers are constantly trying to apply this method,which is more suitable for processing large amounts of data,to hazy space-time analysis.Remote sensing images are indispensable spatio-temporal data in spatio-temporal analysis of haze and convolutional neural networks,due to their special network architecture,making them the most effective methods in image research.This paper studied the spatio-temporal evolution of haze based on the Convolutional Neural Network(CNN)method.The specific steps and contents of the study are as follows:First,the research on the haze classification problem of remote sensing images was first carried out.By using traditional methods to retrieve the remote sensing image from Aerosol Optical Depth(AOD),the data types that can be input as convolutional neural networks were determined.Using the national haze concentration level as the data calibration standard,the constructed convolutional neural network is trained to achieve the classification of haze concentration levels;finally,between the traditional method and the results of this paper and the actual haze concentration level correlation analysis and comparison confirmed the practicality of the convolution neural network haze classification.Further more,in order to predict the haze concentration change,firstly the autocorrelation method was used to analyze the correlation of haze concentration over time,and then the convolution-recurrent neural network structure was built.Using this network,the haze concentration was graded on a time scale.Then,in order to predict the haze concentration level on a finer time scale,this paper build a one-dimensional convolutional neural network method for haze concentration level prediction.In this paper,the Gated Recurrent Unit(GRU)method for comparison was used,highlighting the advantages of one-dimensional convolutional neural network training speed;Finally,using the input of different lengths of time,the law of haze time dimension Conducted research and analysis.Ultimately,in order to study the time-spatial variation of haze,a multi-convolutionjoint network structure based on the You Only Look Once(YOLO)method is built in this paper for the original remote sensing image using partition block processing.Different blocks correspond to different geographical areas,so the output corresponds to the haze concentration levels in different areas.Firstly,mathematical statistics were made for each block output.Then,in order to analyze the data results more precisely,the Moran' I index was used to study and analyze the spatial autocorrelation.Finally,using the Spatial Indicators of Spatial Association(LISA)to study the spatial agglomeration of haze and haze on the time scale,the rule of spatio-temporal evolution of haze in Beijing was obtained.