Spatial-spectral Classification Of Remote Sensing Images Based On Convolutional Neural Network

With the development of imaging spectrometer technique,researchers have higher and higher requirements for remote sensing image processing algorithms.Remote sensing image classification is an important branch of remote sensing image processing.A high-accuracy remote sensing image classification algorithm is conducive to efficiently guiding precision agriculture,mineral exploration,urban planning and other aspects of industrial and agricultural production while saving manpower,material and financial resources.Convolutional neural network is a popular algorithm in the field of deep learning in recent years,and has been introduced into remote sensing image processing by researchers.How to properly design the network structure,adjust hyperparameters,and improve the performance of the network in various specific fields has always been a hot issue in the research of convolutional neural networks.This thesis introduces the application of convolutional neural network in the research of remote sensing image classification from two aspects of multispectral remote sensing scene classification and hyperspectral remote sensing image classification.Various improvement methods of convolutional neural networks are studied from the perspective of enhancement and so on.The research topics in this thesis mainly include the following aspects:(1)Using data augmentation,convolutional block attention module(CBAM)combined with convolutional neural network framework Alex Net,Res Net,Dense Net,VGG to achieve improved classification accuracy.The remote sensing scene classification was carried out on the UC Merced Land Use dataset and the Tiangong-2dataset.The accuracy rate was calculated and the confusion matrix was drawn.Compared with the comparison algorithm,the accuracy rate of the method in this thesis increased by 3.31% and 0.71%,respectively.(2)A weighted feature fusion algorithm based on simulated annealing algorithm is proposed.The existing feature fusion algorithm directly splices or adds multiple features,and the method in this thesis multiplies the features of each branch by a weighting coefficient to splice,and optimizes the weighting coefficient through a simulated annealing algorithm.Based on this,a remote sensing image classification structure based on a three-dimensional convolutional neural network(3D CNN)is designed.Aiming at the problem that ordinary 3D CNN usually extracts features from one scale,some detailed information will be lost,and the performance of Few-Shot Learning is not ideal,this thesis proposes a three-branch 3D CNN,which extracts features from multiple scales.The weighted feature fusion method is used to integrate information from various branches.The method expands the sample size by rotating90°,180°,270°.The method in this thesis uses 10% of the samples for training on the Indian pine data set,the University of Pavia data set,and the Salinas data set,and the classification accuracy of the comparison algorithm is increased by 2.58%,0.30%,and 1.08%,respectively.