Comparative Research On Lidar Equation Inversion Based On Optimized Artificial Neural Network

Atmospheric aerosols can affect solar radiation through scattering and absorption.Atmospheric aerosols play a crucial role in the Earth-atmosphere radiation budget.Atmospheric aerosol particles can be detected by lidar with active remote sensing detection technology.Optical parameters and microphysical properties of various aerosol particles can be inverted by analyzing lidar detection data.Accurately detecting and studying the optical and microphysical properties of atmospheric aerosols has far-reaching significance for revealing the role of atmospheric aerosols in global climate change and atmospheric pollution.Artificial neural network has nonlinear mapping ability and self-learning ability.The optimal solution of lidar function equation can be realized by using artificial neural network.Artificial neural network effectively overcomes the barriers of complex mathematical calculation and many conditional assumptions in solving Mie scattering lidar equations.Therefore,in order to better inversion of atmospheric aerosol extinction coefficient,the paper carried out the research of inversion of atmospheric aerosol extinction coefficient by artificial neural network.The specific work is as follows:Firstly,the Mie scattering lidar system was used to detect aerosol data,and the original data sets of three wavelengths(355nm,532nm,1064nm)were obtained under different weather conditions.The original detection data were preprocessed and normalized.In order to avoid the influence of non-geometric overlapping region data on the experiment,data with a height range of 2?12km were selected for the experiment.Secondly,the BP neural network constructed in this paper was used to carry out the inversion experiment of atmospheric aerosol extinction coefficient.Aiming at the problems of BP neural network easily falling into local minimum,low experimental accuracy and slow convergence,the model of BP neural network was optimized immediately.Elman neural network with local memory unit and local feedback and Wavelet neural network with time-frequency domain and zoom were constructed by adding feedback layer and transform transmission function.The extinction coefficient was inverted by Elman neural network and Wavelet neural network.The mean square error values of the experimental results of the two networks were 10^-14?10^-13 and 10^-13?10^-12 orders of magnitude,respectively,and the unit was(km^-1)²,which improved the accuracy of the experimental inversion results and reduces the experimental error.It was proved that the overall accuracy and performance of the experiment using the optimized neural network were further improved.Finally,the lidar detection data with wavelength of 355nm,532nm and 1064nm in foggy weather and the mixed cross sample set were used to verify the generalization ability and applicability of the two optimized neural network models.The experimental proof has been formed in the three wavelengths of 355nm,532nm and 1064nm,and in the light haze and fog weather conditions in the longitudinal direction.In foggy weather,the mean square error of the two neural networks has increased,but it is still between10^-14?10^-13 and 10^-13?10^-12orders of magnitude.After the training of mixed cross sample sets,the mean square error values were significantly reduced between 10^-15 and10^-14.The reliability and applicability of two optimized neural networks to invert extinction coefficients were proved.In addition,the specific value of Elman network inversion error was lower than that of Wavelet network,which further proved that the effect of Elman neural network inversion extinction coefficient was better than that of Wavelet neural network.