Hyperspectral Remote Sensing Image Unmixing And Object Detection Based On Sparse Autoencoder

Hyperspectral remote sensing technology uses imaging spectrometer to obtain hyperspectral images containing geometric and spectral information of ground objects,which is widely used in military reconnaissance,public security,pollution monitoring,soil investigation,precision agriculture and other fields.Therefore,it is necessary to develop efficient hyperspectral data processing technology.There are a lot of mixed pixels in hyperspectral image.This process of decomposing the mixed pixels into several kinds of single object pixels and obtaining the proportion of each single pixel in the mixed pixel is called hyperspectral unmixing.The process of dividing the measured pixel into target pixel and background pixel according to the spectral characteristics is called target detection.Hyperspectral image has the characteristics of multi band,strong correlation between bands,data redundancy and high resolution,so there are many technical difficulties.How to fully and effectively mine the information contained in hyperspectral images is a problem faced by hyperspectral unmixing and target detection.In essence,hyperspectral unmixing and target detection are processed according to features.Sparse self-encoder is an effective tool to extract features.Wavelet transform can extract multi-scale features of signals.So in this paper,we construct sparse unmixing network of wavelet multi-scale feature fusion auto-encoder,mine the deep features of hyperspectral image data,extract endmembers,carry out abundance inversion and apply the network output layer data with rich background information to the hyperspectral target detection problem.In this paper,we propose a sparse auto-encoder based algorithm for hyperspectral remote sensing image unmixing and target detection.The main work is as follows:(1)A hyperspectral unmixing algorithm based on wavelet multi-scale feature fusion auto-encoder is proposed.By using the deep learning auto-encoder and the multi-scale feature of wavelet domain,the endmember extraction is regarded as a feature learning problem.By modifying the network structure and loss function,the learning ability of the auto-encoder network is improved,the endmember is extracted and the abundance is estimated.The results of simulation and real datasets show that this method can extract the endmember and estimate the abundance of hyperspectral images more accurately,and get better results of unmixing.(2)In order to separate the complex background information from the target information,an anomaly detection algorithm based on sparse auto-encoder is proposed.Firstly,the hyperspectral image to be measured is de-mixed with sparse auto-encoder,and the output data of encoder layer contains rich background information,so we consider introducing spectral unmixing technology into anomaly detection,building reconstruction error,effectively restraining the interference of background to anomaly detection,and improving the accuracy of anomaly detection.Experiments on both simulated and real datasets show that the anomaly detection results of this method are better than those of other mainstream algorithms mentioned in this paper.(3)A prototype system of hyperspectral image unmixing and target detection based on sparse auto-encoder is designed and implemented.The system integrates a variety of unmixing and target detection algorithms,and provides a relatively complete hyperspectral unmixing and target detection process: from inputting the hyperspectral image,selecting the unmixing algorithm,saving and outputting the endmember and abundance map,selecting the target detection algorithm,and finally outputting the target detection results,which has a good prospect of engineering application.