Study On Parallel Algorithm Of Aerosol Optical Depth Retrieval From Remote Sensing Image Based On Geostationary Satellite Himawari-8

In recent years,the frequent occurrence of hazy weather in central and eastern China has attracted widespread public attention.The culprit of haze is fine particulate matter(PM_2.5).Daily measurement by the environmental protection department can provide accurate PM_2.5data,but this routine measurement can only be performed on limited ground sites,and there are fewer sites in rural and western China.In contrast,satellite remote sensing has the characteristics of wide coverage and clear macroscopic changes.It can fully reflect the distribution and transmission features of particle in large aero scale.Aerosol Optical Depth?AOD?is one of the important optical properties of atmospheric aerosols,and is also the main parameter for estimating PM_2.5 concentration by remote sensing.Compared with conventional polar-orbiting satellite which can only achieve 1-2 observations per day in the same region,geostationary satellite have the outstanding advantage of high time resolution which is important for the research area of cross-border transmission of pollutants and analysis of haze formation.However,geostationary satellites have fewer bands and larger data,there are still many problems in aerosol retrieval accuracy and real-time,which can't meet the business needs of environmental protection meteorological deparments in China.In order to produce aerosol optical thickness remote sensing products with all-weather,high-precision and real-time,this paper presents three solutions correspondingly in the aspects of retrieval algorithm and real-time processing based on Himawari-8 satellite band data.1.Aerosol model construction.In satellite observation,the establishment of aerosol model is through clustering ananlysis base on global long-term sequence ground-site data.However,in some certain aeres and countries,aerosol model on the global scale may be not applied.Especially for areas or countries with serious atmospheric particulate pollution,aerosol components are complex and strong variability,the suitable aerosol model to construct lookup table is crucial.This article will use the five aerosol models in VIIRS-AOD-ATBD and select the most proper aerosol optical thickness as the observation value of the Himawari-8 through threshold's judgement and constraint.2.Surface reflectance construction.For the Dark Target algorithm,the accurate surface reflectance relationship to achieve surface-atmosphere decoupling is the most important issue for Aerosol Optical Depth retrieval.Regarding to different surface types,the difference in surface reflectance between visible and near-infrared channel is huge.In this paper,the surface reflectance obtained by atmospheric correction is used to perform regression analysis and training based on normalized vegetation index to get the best fitting relationship and its NDVI classification threshold between each band.And finally the reflectance relationship useful for Himawari-8's 0.46?m-0.64?m and 0.64?m-2.3?m band is built successfully.3.AOD retrieval performance.In the process of aerosol optical thickness retrieval,there are two bottlenecks of the overall performance.First is the computation to call radiation transmission model to construct lookup table through setting aerosol model,and the second aspect is to invoke lookup table to retrieve AOD.With the continuous development of GPU general-purpose computating,its powerful floating-point computing capabilities,high-density computing methods,and cost-effective advantages provide an effective solution to the above problems.Based on CPU and GPU's parallel collaborative processing of CUDA heterogeneous computing system,this article will programme retrieval task rationally and make full use of the respective advantages of CPU and GPU,and then realize the acceleration of AOD retrieval algorithm of Himawari-8.The correlation coefficient between AOD of retrieval algotithm and AERONET observations is up to 0.87,the regression equation is=0.93+0.04.Besides,the correlation coefficient in different seasons is also good enough.Comparing with the Terra,Aqua,and Himawari-8 standard AOD products in regional scale,the retrieval results present consistence in distribution and accuracy as well.In the CUDA parallelization process,the algorithm is optimized from three aspects:task allocation,thread setting,and storage mode.The test results show that the performance of the parallel algorithm is obviously improved,and with the increase of the image size,the acceleration ratio is continuously improved and gradually stabilized at about 5.5.