Aerosol Optical Thickness Retrieval Based On Neural Networks

Atmospheric aerosol is a multiphase system consists of solid particles, liquidparticles and gas carrier suspended in the atmosphere. It has an important impact onhuman life, atmospheric environment and global climate change.Aerosol OpticalThickness (AOT) is one of the most critical parameters to characterization aerosol，and is also a key factor to determine the climate effect of aerosol. China has a vastarea, complicated surface condition and diverse ecological types. As the ground site oflong-term observations for aerosol is less in China， using satellite data to retrieveAOT has become an important supplementary means to obtain the regionaldistribution of the aerosol. How to retrieve AOT more quickly and accurately hasbeen the difficulty and hotspots in international research.This paper is the study of the AOT retrieval method based on the Landsat-7satellite Enhanced Thematioc Mapper Plus data(ETM+).The key point of the retrievalof aerosol optical thickness over land is how to remove the ground contribution fromthe information received in satellite.In this paper， use apparent reflectance innear-infrared to determine the dark pixel， then calculate the ground reflectance withthe relationship between the2.1um band and the red-blue band in DDV.The sourcecode of6S radiative transfer model was to be rewritten with OpenMp， to achieveparallel computation of6S model.To complete the construction of the lookup tablerapidly by inetwork in inputing a separate parameter to the model， accomplish theconstruction of a neural network by training the lookup table with Matlab neuralnetwork toolbox.The retrieval accuracy and efficiency of the two algorithms were compared andverified.In aspect of the accuracy of the algorithm， taking into account there is alarge different spatial scales between the retrieval results of ETM+data and MODISaerosol optical thickness products，we resampling the retrieval results of the the twoalgorithm to10KM spatial resolution，then take comparison with the MODIS aerosoloptical thickness products in space.In addition，we also utilize the data of AERONETto validate the accuracy of the two algorithms.In aspect of the retrieval efficiency，we take comparison of the needed time of the two algorithms in discrete number ofpixels.The results showed that it can improve the retrieval efficiency， shorten theretrieval time required with using the neural network， especially for massive amountof data， can greatly enhance the efficiency of the retrieval.