Neural Network-based Aerosol Remote Sensing From Multi-angle Polarimetric Measurements Over Ocean

The ocean accounts for 71% of the world’s area and is also the Earth’s largest carbon reservoir.Aerosol remote sensing over the ocean will help improve the quantification accuracy of water color remote sensing,quantifying marine biomass,estimating ocean primary productivity,and even global climate change.With the development of remote sensing detection technology,the SPEXone onboard the PACE satellite that NASA is about to launch,and the DPC onboard the Chinese GF5(02)are all multi-angle polarization sensor,which will provide abundant data for the highprecision inversion of aerosol properties over the ocean.However,the optimaized real time inversion algorithm based on the verctor radiation transfer model and traditional look-up table algorithm have certain limitations and are not conducive to satellite operational data processing.This research has developed an aerosol inversion algorithms over the ocean based on neural network models and applied to different regions,space-borne,airborne,simulated and measured data.The combination of artificial intelligence and physical models in the inversion of aerosol over the ocean with multi-angle polarization measurements has been realized.Under the premise of ensuring the accuracy of inversion,it greatly improves the efficiency of satellite remote sensing operational calculations.This research has conducted the following research:1)Introduced the current development history of polarization sensors used in aerosol research at home and abroad,and the mainstream algorithms for aerosol polarization inversion over the ocean.Although the method based on the radiation transmission model is accurate in calculation,it is very time-consuming and cannot meet the needs of operational data processing.The method based on the look-up table can improve the calculation efficiency to a certain extent,but as the polarization sensor moves towards hyperspectral multi-angle.It also poses new challenges to the calculation and storage of look-up tables.Artificial intelligence has great advantages in massive data processing,which is conducive to aerosol satellite remote sensing over the ocean and its operational data processing.Therefore,this study proposes a research idea of combining artificial intelligence and physical models of multi-angle polarized load aerosol retrieval over the ocean.2)Developed an aerosol retrieval algorithm over the ocean based on neural network model.This research analyzes the physical process of radiation transfer in the atmosphere–ocean coupling system.The atmosphere model considers a multi-modes aerosol model which mixes coarse and fine mode particles.In the terms of ocean models,it is based on a single-parameter chlorophyll concentration ocean bio-optical physical model for ocean water bodies.Then the neural network topology design and model training are carried out.The input parameters are determined as the chlorophyll concentration,incident angle and outgoing angle,and the output parameters are the first100 components of the ocean optical characteristics that can be reconstructed all wavelength information after principal component analysis.Through a large number of comparative verifications,it is finally determined that the neural network topology is 3,20,30,20,100 and the verification accuracy of the model can meet the actual needs of aerosol inversion.3)In this study,the linearized vector radiation transfer model LINTRAN was used to solve the radiation transfer equation based on the Gauss-Newton iterative algorithm of Phillips-Tikhonov regularization,and the inversion of different types of data were carried out and validated.For the SPEXone space borne simulation data,the inversion of aerosol characteristics can basically meet the design accuracy.For the inversion of the SPEX airborne measured data in the ACEPOL campaign,the results have been compared with the lidar HSRL data and the ground-based AERONET site data with a high consistency.For the Gaofen5 DPC satellite data,the inversion was carried out for a whole month and verified with ground-based stations distributed around the world.The root mean square error of the aerosol optical depth is 0.16 and the average relative error is 0.115.In summary,this research has realized the combination of artificial intelligence and physical models for aerosol retrieval,developed a neural network-based aerosol inversion algorithm over the ocean,and performed inversion for different regions and different types of data.The results show that the algorithm has high inversion accuracy,strong universality,and greatly improves the inversion efficiency of aerosol retrieval over the ocean.It will provide favorable support and guarantee for the operational processing of aerosol characteristics over the ocean for multi-band,multi-angle polarization loads,such as SPEXone,DPC,etc.,which will be launched in the future.