Compression Based Deep-learning Assisted Analysis Of Polarimetric Scattering From Cylindrical Components Of Vegetation

Vegetation plays a very important role in the earth's ecosystem and is known as the lung of the earth.It is closely related to the carbon cycle and water cycle.Remote sensing technology,based on the study of electromagnetic scattering characteristics of vegetation,can obtain important information such as biomass and water content in vegetation area,which is of great significance for guiding crop planting and environmental monitoring.In the study of the electromagnetic scattering characteristics of vegetation based on the discrete medium model,finite cylinders are often used to replace the trunk and branches of vegetation.Due to the complexity of ground objects,an efficient scattering modeling method for finite cylinders is very important.With strong feature extraction capability,neural network can deeply learn the higher-order function of the target model based on the training data set,and it has high computational efficiency after the completion of training,only involving a single forward operation.Based on the deep neural network,this paper explores the scattering characteristics of vegetation cylinder components from the data level.The main work and achievements are as follows:1.Deep Neural Network with Discrete Cosine Transformation(DNN-DCT)was used to capture the scattering pattern of a finite length cylinder.Based on the discrete cosine transform(DCT)to eliminate the redundancy of the scattering data from a finite cylinder,a DNN-DCT model is proposed to learn the scattering amplitude function from a finite cylinder.DNN-DCT can quickly learn tens of millions of orders of magnitude data and complete the capture of scattering patterns in multi-parameter dimensions.Numerical results show that the converging model is robust to the interpolation results of either single variable or multiple variables,and the physical reciprocity theorem is also satisfied.Meanwhile,compared with the T matrix method based on virtual segmentation,the DNN-DCT model has a 4-6 order of magnitude increase in computational efficiency.2.Application of depth compression sensing technology in electromagnetic scattering analysis of finite length cylinder.In this paper,the depth compression sensing technique is introduced to reconstruct the scattering wave data of a finite cylinder irradiated by a wide range of incident angles with a few observations.Compared with the traditional compressed sensing system,which adopts hand-customized sampling matrix,sparse base and complex convex optimization and greedy algorithm,the deep compressed sensing framework in this paper learns the nonlinear representation and sampling matrix from the training data.After the training,the model can achieve lower sampling ratio and faster reconstruction speed compared with the traditional compressed sensing technology.