Fast Calculation Of Electromagnetic Scattering From Rough Surface And Study Of Parameter Inversion Based On CNN

Research on electromagnetic scattering properties of randomly rough surface provides theoretical foundation for target detection,recognition and radar precise guidance,and has also been applied in agriculture,marine fishery and other civil fields.In this thesis,I present a fast computational method for composite electromagnetic scattering from multi-targets above a randomly rough surface,a fast analysis method for wideband responses of a rough surface,and an inversion method of rough surface parameters using deep convolutional neural networks.The thesis is organized as follows:1.A new scalar integral equation-based single-source equivalence principle algorithm is proposed,which is applied to calculate composite scatterings from multiple PEC or dielectric objects above a dielectric rough surface.The rough surface and objects are divided into different regions by equivalent surfaces.The unknown density on the equivalent surface is much less than those on the scatters,and the equation of SS-EPA only contains electric currents(or magnetic currents)using the relationship between electric and magnetic currents on the equivalent surface.Therefore,SS-EPA can significantly reduce the unknown numbers of the MOM method.In the numerical experiments,composite scatterings from different numbers and types(PEC or dielectric)of objects above a Gaussian dielectric rough surface are simulated.The results show the proposed method exhibits higher computational efficiency when more objects with more complex shapes are simulated.2.A hybrid method combining Maehly approximation with phase extraction technique is proposed to solve the wideband electromagnetic scattering problem of a rough surface.Instead of the traditional MOM method,in order to further accelerate the solution,a CBF-ACA hybrid method is proposed to solve the electromagnetic currents at frequency points corresponding to the Chebyshev nodes.The present method can be applied not only to Gaussian rough surfaces with different root mean heights,but also to sea surface with different wind speeds.The numerical results show that the proposed algorithm can improve the speed of computation with good applicability and stability.3.A novel inversion method for rough surface parameters using deep convolutional neural networks is proposed.By training of the limited number of SAR images of rough surfaces,the parameters of any rough surface within the given parameter range can be predicted.The CNN network present in the thesis is constructed on the basis of simulated data,which provides a new idea for the training of deep convolutional neural networks.The RMS heights and correlation lengths of rough surfaces can be inverted simultaneously using the proposed CNN network.Numerical simulations show that the inversion results of the CNN network are better than those of the SVM method.