Research On Anomaly Detection Algorithm Of Hyperspectral Image

Hyperspectral image is one kind of remote sensing image,which can obtain rich spectral information as well as spatial geometric information of land covers.The spectral dimension of a hyperspectral image consists hundreds of continuous narrow bands in the wavelengths ranging from visible to infrared.Therefore,hyperspectral image offer a unique ability to capture subtle physicochemical properties of land covers.Hyperspectral anomaly detection is an important branch of hyperspectral image processing.Compared with target matching detection requiring some prior information about the anomaly and background,hyperspectral anomaly detection can automatically detect some pixels or regions which are quite different from local or global background distribution,providing some regions of interest for further interpretation.Hyperspectral anomaly detection has been widely used in military and civilian fields.Although researchers have put forward some effective theories and methods,there are still some deficiencies and problems: 1)The loss of local geometric structure information in the process of background modeling;2)The background modeling may be contaminated by some anomaly pixels;3)The discriminant anomaly degree which separate anomaly and background clearly.Hence,our paper will focus on the above three key issues and try to propose some hyperspectral anomaly detection methods.The main research contents are as follows:1)To address the loss of local geometrical information in the process of background modeling,a hyperspectral anomaly detection algorithm based on locality constrained low rank representation and dictionary learning is proposed.For this method,we introduce a locality constrained term to preserve the local geometrical structure.Motivated by the idea of manifold learning,an affinity matrix is used to constrain the representation coefficient matrix,in order to force the pixels with small distance to have similar low representation vector.Moreover,the dictionary learning is integrated into the low rank representation model,and the dictionary will be updated iteratively along with the optimization process of LRR.In the end,the residual matrix is used to calculate the reconstruction error of pixels to detect the anomaly targets.The visualized and quantitative results demonstrate that the locality constrained term can greatly reduce the loss of local structure information in the process of background modeling by low-rank representation.2)To cope with the loss of local geometry information in the process of background modeling,a hyperspectral anomaly detection algorithm based on local structure preserved autoencoder is proposed.First,a manifold learning method,diffusion mapping,is used to obtain a low-dimensional embedding of the original data.Then,the embedding is used to guide the training process of the encoder,so that the low-dimensional code extracted by the encoder contain both structural and spectral information.In addition,a locally smooth regularization term is introduced into the objective function of the encoder,in order to maintain the local structure between the training pixels.The visualized detection results verify that our technical routes can preserve the local structure information in the process of background modeling with autoencoder.3)To solve the problem that the background modeling may be contaminated by anomaly pixels,a background purification framework with extended attribute profile is proposed.Considering that the anomaly objects are isolated,closed and small-size,the connected components conforming to the spatial attributes of anomaly objects are extracted by using the extended attribute profile,and then the connected regions are removed from original image to obtain pure background dataset.Further,the proposed framework is applied to three different hyperspectral anomaly detection algorithm: the RX algorithm,the low rank and sparse representation-based detector and autoencoderbased detector.By improving the background data acquisition of aforementioned algorithms,the proposed framework can enhance the accuracy of their background model.The visualization results and quantitative evaluation indexes shows that the framework could solve the problem of anomaly target pollution in the process of background modeling.4)To address the anomaly degree measure,a hyperspectral anomaly detection method based on the data mass estimation is proposed.Considering that the anomaly objects are isolated,sparse and having large spectral difference,we construct an ensemble of half space trees to estimate the data mass of every pixels in a hyperspectral image,in order to obtain the anomaly degree of pixels.Considering how isolated an pixel is from its local neighborhood,and local anomalies may be masked by the background clusters of low density,We borrow the concept of relative density and further bring forward an improved metric called the relative data mass.Compared with other methods requiring background modeling,our method measures the anomaly degree from the perspective of its own particularity.The visualization results and evaluation indexes show that the proposed method can effectively separate the anomaly pixels from the background.