A Deep Learning-based Target Detection Algorithm For Hyperspectral Remote Sensing Images

Hyperspectral remote sensing images contain rich spatial dimension information and spectral dimension information,so they can distinguish subtle spectral differences on extremely similar materials,thus solving many problems that cannot be solved by multi-color images and multi-spectral images.It is widely used in the field of target detection and can be widely applied to military,agricultural,mining,and so on.Deep learning is a new field in machine learning research.It interprets data by establishing a neural network that simulates human brain analysis learning.Therefore,some scholars apply deep learning to hyperspectral fields,and Convolutional neural networks are the most widely used network in the field of deep learning.Therefore,based on the deep learning model of convolutional neural network,this paper studies the target detection of hyperspectral remote sensing images.The main work is:(1)At present,the hyperspectral remote sensing image target detection model based on convolutional neural network only considers the spatial features of remote sensing images,and can not fully extract the spatial-spectral features of hyperspectral remote sensing images.However,because hyperspectral remote sensing image data is rich in spectral information,it can achieve good detection results in the target detection field.Therefore,based on 1D-SSCFECNN,2D-SSCFECNN and SSCFERCNN,three kinds of hyperspectral remote sensing image target detection models are proposed in this paper.The model is introduced in detail for the process of spatial-spectral feature extraction.The advantages of these three new models compared with the existing hyperspectral remote sensing image target detection models based on convolution neural network are analyzed,and target detection experiments are carried out on real data sets and synthetic data sets respectively.The experimental results show that compared with the classical target detection algorithm and the existing target detection algorithm based on deep learning,the proposed three hyperspectral remote sensing image detection model of spatial-spectral features extraction based on 1D-SSCFECNN,2D-SSCFECNN and SSCFERCNN can achieve better detection accuracy.Finally,this paper analyzes the advantages and disadvantages of these three algorithm models,and summarizes the suitable application scenarios.(2)The deep learning model generally takes a lot of time to adjust the parameters and training network.This paper takes the 1D-SSCFECNN model based hyperspectral remote sensing image target detection as an example,and analyzes its very cumbersome parameter adjustment process and results.There are some correlations between many hyperspectral remote sensing image data in real life.Therefore,this paper analyzes the feature transfer learning algorithm based on convolutional neural network applied to the field of hyperspectral image target detection by analyzing the correlation between different data sets,proposes a feature transfer algorithm based on the transfer matric of average value of the category as to these problem,and implements hyperspectral remote sensing image target detection method with correlative data set.Finally,Pavia C hyperspectral data set is used as target domain data set and Pavia U hyperspectral data set is used as source domain data set to carry out experiments,and the corresponding target detection accuracy and time are given.Compared with the accuracy of training target data alone,the target detection model of hyperspectral remote sensing image based on TMAVC feature migration algorithm can greatly save the time of target detection without sacrificing the accuracy of target detection.