Research On Electromagnetic Scattering Mechanism Of Multi-configuration Electric Large Urban Targets And Urban Area Classification

Synthetic Aperture Radar(SAR)system is an efficient tool to monitor the earth by transmitting and receiving electromagnetic wave.With more and more SAR systems launched,Oceans of data were acquired.The innate character of target underlying these data can be revealed by understanding the target scattering mechanism.The electric-large buildings in urban area havemultiple configurations which lead to complex scattering mechanisms.The knowledge of urban scattering mechanisms has not met the requirement of civil and military applications yet.For this purpose,we commit ourselves to investigate two basic issues:scattering mechanism and classification The main content of this dissertation can be summarized as the following four parts:In the first part,we proposed the scattering model of urban buildings.Franceschetti treats the scattering model of isolated building as the compound scattering of parallelpiped and random rough ground.Then the analytical expression is derived by combining geometrical optics(GO)and physical optics(PO).Based on this,we derived the analytical expression by combining GO and small perturbation approximation(SPA)when the random rough ground does not meet the condition of GO or PO.The scattering of building is numerically simulated by different azimuth angles,incident angles and parameters of ground.Furthermore,we give the analytical expression of even scattering of couple of buildings by analyzing the equivalent height with the geometrical model of buildings and incident angle.It's very difficult to acquire the analytical expression of scattering from multiple buildings.We use ray tracing method to simulate the backscattering of multiple buildings in this dissertation.In the second part,a model decomposition method for urban is proposed.Freeman decomposed the scattering into three parts:odd scattering,double scattering and volume scattering.But this decomposition method would overestimate the volume scattering part.Yamaguchi adds a helix scattering component to Freeman's model and a branch condition to choose a suitable volume scattering model.The volume scattering part is reduced by Yamaguchi four-component decomposition(YO4).YO4 together with polarimetric orientation angle(POA)compensation method would further reduce the volume scattering by minimizing the cross-polarized channel.Also four-component decomposition method enhances the double scattering in urban area with small azimuth angle,but it becomes invalid in the area with large azimuth angle.Moreover,both Freeman's three component decomposition and Yamaguchi's four component decomposition produce negative power.To solve the two problems above,we proposed an adaptive oriented buildings scattering model decomposition(AOBSMD)method.First,the relationship between orientation angle of buildings and POAs is analyzed.Then the adaptive oriented building scattering model(AOBSM)is built based on the relationship.At last,the AOBSMD is constructed by combining AOBSM and the four scattering model of YO4.An optimization strategy which eliminates the negative power is proposed to estimate the scattering contributions.The real SAR data confirms that the proposed method can reveal the true scattering property of buildings with large azimuth angle.In the third part,we study the urban classification methods based on the scattering features of targets.This part consists of three perspectives:(1)classification method by combining statistical distribution and scattering features.First we use the ABOSMD result to get an initial segmentation.Under the assumption of the polarimetric SAR data meeting Wishart distribution,the segmentation map can be obtained by an iterative clustering process.Even though this is an unsupervised classification method,but we can still label the ground type based on its scattering features.(2)Classification method based on scattering features selection.The polarimetric decomposition features are sensitive to the ground targets.The importance of each feature is evaluated by random forest(RF)method.The features are sorted in descending order.Then a sequential forward selection(SFS)algorithm is used to optimize the feature set.The data can be classified by the optimized feature set.The real SAR data proves that the proposed method can achieve higher performance by fewer features.(3)Urban area classification with polarimetric statistical features of simulated data.The scattering property of targets with different orientation angles is simulated by rotating the coherency scattering matrix with a series of angles.Polarimetric features would be extracted in each rotation angle.Then a polarimetric statistical feature vector would be acquired for urban classification.The real SAR data demonstrate that the proposed method can work efficient in urban area with large azimuth angles.In the fourth part,we study the deep learning models in urban area classification.The deep learning models representing as convolutional neural network(CNN)has achieved high classification accuracy of PolSAR data.This dissertation cooperate deep learning models with scattering mechanism of target to improve the performance of urban area classification.(1)Because the model decomposition method is very effective in PolSAR classification.We propose a polarimetric feature driven CNN model for urban classification by combining CNN and model decomposition features.This model can achieve high performance by very limited training samples.(2)Two sequences of target scattering data with the independent variable of rotation angles can be obtained by rotating the coherency matrix with two different SU(3)matrixes.Three-dimension convolutional layer is used to extract the space information and the LSTM is used to extract the sequence information.The urban is classified by bagging the features extracted by two sequences of data.The proposed method has the characteristics of fast convergence,high classification accuracy and very limited training samples.