Research On Spectral-spatial Information-based Target Detection Algorithms For Hyperspectral Images

As an image cube,hyperspectral remote sensing data contains much richer feature information than panchromatic or multispectral images,and has huge development potential.The target detection algorithm of hyperspectral images(HSI)is one of the hottest topics in this filed for decades.However,whether in the mainstream detection algorithm based on spatial information in image processing or the detection algorithm based on spectral information derived from traditional signal theory,only a small part of the information in HSI is used,while a large amount of useful information has not been exploited.The lack of labeled samples in the field of hyperspectral remote sensing also restricts the application of emerging popular algorithms such as deep learning,which hinders the development of target detection algorithms.Therefore,it is of great practical significance to explore and develop adaptive hyperspectral target detection algorithms that fully integrate the spatial spectrum information based on different prior information conditions.The research of This paper is as follows.In the case of lacking targets and background prior information,the paper first studies the sparse representation model of hyperspectral images and the problem of solving the sparse coefficients,and uses non-negative sparsity score estimation algorithm for unsupervised anomaly target detection.The algorithm bases on the spectral sparse characteristics of anomaly target and the utilization rate of sparse dictionary atoms.Post-processing using spatial filtering based on conditional random field is combined to make full use of the pixel label and spectral similarity among the pixel neighborhood to further refine the detection result.Through the combine of spectrum and spatial information,the detection performance of this algorithm is improved.For the detection of hyperspectral targets with a small number of target prior spectral samples,this paper proposes a target detection framework based on a 1D siamese convolutional neural network(1D SCNN).1D SCNN is used to extract the spectral features,and then the Euclidean distance in the feature space is used to fit the similarity of the input spectrum and as a measure for the matching of target pixels.The siamese network is suitable for the case of small samples,which can use pixel-pairs to train effectively so as to solve the problem of insufficient training samples in HSI.In addition,based on the preliminary detection results with spectral information,we use spatial filtering to combine context information in order to strengthen the correlation between pixels,which could effectively improve the performance of the algorithm.Finally,with certain target and background label prior information,this paper proposes a target detection algorithm based on multi-scale fully convolutional neural network(MSFCN).Considering that there is a lot of redundancy in the spectral dimension of HSI,we introduce the principal component transformation to preprocess the data and uses the corresponding multi-scale 3D convolution kernels to fully extract the spatial-spectral feature among the high spatial resolution HSI.At the same time,aiming at the scarcity of training samples for HSI,a data augmentation method combined with spatial geometric transformation and spectral probability statistical model is used to expand the training sample set in order to avoid the overfitting of the model.Comparative experiments show that this HIS target detection algorithm achieves outstanding reliability and effectiveness.