Research On Anomaly Detection Method Of Hyperspectral Remote Sensing Image

Anomaly Detection(AD)in hyperspectral image(HSI)data has received extensive attention in various applications.The target of anomaly detection is to detect pixels in a hyperspectral data cube which spectra differ observably from the background spectrum.Anomaly detection is an unsupervised object detection problem that aims to identify anomalous components in data,which has always been a challenging problem due to the complexity of hyperspectral image backgrounds and the limited number of object samples.Anomaly detection can be modeled as an unsupervised binary classification problem between background and anomaly classes.The challenge of this problem is that there is no prior knowledge about anomalies or the background,which may have complex textures and will increase the difficulty of detection.Furthermore,HSI data is often affected by noise due to the limitations of HSI acquisition equipment.But there are two important features that can be used to distinguish an anomaly from its background: 1)the anomaly is sparse compared to other objects;2)the anomaly has unique characteristics compared to the surrounding background.Aiming at the above-mentioned problems in anomaly detection of hyperspectral remote sensing images,this paper introduces a deep learning method,studies the use of the advantages of neural networks to extract deep features of hyperspectral images,and combines a variety of background suppression methods to reconstruct a more discriminative background.The main research works of this paper contain the following contents:1)Aiming at the low accuracy of hyperspectral anomaly detection in complex backgrounds,a hyperspectral anomaly detection method based on joint clustering and autoencoder learning is proposed.First,the original hyperspectral image is dimensionally reduced using an autoencoder with mean-shift clustering loss,and a dictionary is constructed from the hidden layer data after network training.Second,a low-rank decomposition is performed on the constructed dictionary to obtain a sparse anomaly score matrix.Experiments suggest that the scheme can make full use of the more discriminative deep features extracted while reducing the data dimension,and build a dictionary that can better characterize the hyperspectral images of complex backgrounds.Spectral anomaly detection has a positive effect on fully extracting background features.At the same time,in hyperspectral anomaly detection(HAD),due to the lack of prior knowledge,complex imaging environment,low spatial resolution and other problems,it is difficult to label accurately manually.This unsupervised learning method overcomes the dilemma of limited detection performance of label free detection.2)Aiming at the problem of insufficient background learning ability due to limited target samples,a semi-supervised method is designed to improve the adversarial autoencoders(AAE)for anormal detection of hyperspectral images.An anomaly suppression constraint function based on spectral information matching is added to the loss function of AAE.First,an adversarial loss that matches the spatial distribution of the hidden layer with the defined mixture of gaussian prior distributions;second,an anomaly suppression constraint loss is used to ensure that the network is only trained on background samples;then,an autoencoder loss is used to calculate the input samples and reconstruct the output deviation.Finally,anomalous objects are identified by detecting the difference between the reconstructed hyperspectral image and the original hyperspectral image.Experiments suggest that the model can utilize anomaly suppression to constrain undesired anomalous spectra and provide accurate background descriptions.