Research On Automatic Classification Of Typical Features Based On SAR/InSAR Images

The typical feature classification algorithm of SAR/InS AR image is an important branch of SAR image understanding and interpretation,which has important application value.In recent years,SAR feature classification based on deep learning has become the mainstream instead of traditional algorithms.However;the "end-to-end" characteristics of deep neural network make it difficult for people to understand its working mechanism and decisionmaking basis.This "black box" characteristic seriously hinders its performance improvement and wide application in practice.This paper uses the depth learning method to study the automatic classification of SAR/InS AR typical features,and makes a preliminary study on the"black box" characteristics of the depth network.The main research contents and contributions are as follows:1)A network framework of Multi-layer Feature Fusion Attention Mechanism(MF2AM)is proposed.The network uses ResNet101 for preliminary feature extraction and outputs 4level resolution features,and then deeply extracts the multi-scale features of the high-level feature map through Attention-based Atrous Spatial Pyramid(AASP);The constructed Semantic Embedding Branch(SEB)integrates the middle-level features,and then aggregates the spatial information to guide the pixel classification of the feature map and improve the effect of edge segmentation.Experiments show that the network can extract high-precision overlapping targets in large-scale millimeter wave InSAR images.2)Aiming at the problems of regional missing detection and poor effect of boundary detection in surface water extraction of the Yellow River from SAR images,a Global Attention and Boundary Refinement Network(GABRN)is proposed.GABRN extracts multi-scale features from the global range through the Multi-scale Chanel and Space Pyramid(MSCSP);The Fusion Refinement Module(FRM)is used to fuse the low-level features containing boundary detail information and the high-level features containing semantic information,and then subdivide the boundary.Five years of sentinel-1 data experiments show that the network can achieve high-precision water extraction.3)Aiming at the "black box" problem of depth network,an automatic detection framework of surface water in SAR image based on interpretable depth learning is proposed.The framework mainly includes three parts:DeeplabV3+ water extraction network with four different backbone networks,Local and global mixed attribution(LGMA)for backbone network selection,and semantic specific class activation mapping(SSAM)for the visualization of output layer characteristic thermodynamic map.The framework integrates LGMA to analyze the attribution of four backbone networks,and evaluates the feature extraction performance of deep networks through visual feature heat map.The experiments show the effectiveness of the interpretative method for backbone networks in this paper.In this paper,the classification of typical features in SAR/InSAR images is studied.The proposed MF2AM network has obvious performance advantages and realizes the highprecision automatic extraction of overlapping regions in InSAR images;The problem of extracting the boundary of the Yellow River is solved by GABRN,and the detection results of GABRN are satisfactory;The proposed interpretable deep learning framework effectively explains the performance of different backbone networks in SAR image water extraction,and can select the best backbone network for SAR image water extraction..