Research On The Key Techniques Of Hyperspectral Images Classification With Multi-domain Feature Extraction

Originated from the science of remote sensing and promoted by the advancements of the spectral imaging technology and the science of the optical imaging instruments, hyperspectral imaging (HSI) technology has become a promising and rapidly developing technology in image processing and analysis. HSI integrates the advanced spectral technology and the traditional imaging technology in two-dimensional space together, and uses the imaging spectrometer to image in the spectral range which covers dozen or even hundreds of spectral bands continuously distributed, acquiring the spatial features of objects and collecting the continuously varied spectral information, synchronously. And due to the superiorities on the composition analysis, object classification and target recognition etc., HSI has become an advanced technology in many fields, such as geological science, agricultural production, food security and biochemical science etc.Feature extraction and classification technology are always the key point of the HSI processing and applications. Studies and innovations in HSI technology are full of the theoretical significance and practical values. However, the HSI has several disadvantages with large amount of data, more bands with strong correlation, high data redundancy etc., which seriously hinder the developments and applications of HSI technology. This paper focuses on that "how to extract spectral and spatial information from the huge amount of HSI data and classify effectively to improve the classification accuracy and efficiency" and studies the key multi-domain united feature extraction and classification techniques in HSI. The main contributions of this paper can be concluded as follows:Firstly, considering the problems of HSI with huge amount of data, more spectral bands with high correlation, high data redundancy etc., this paper has researched the current feature extraction measures in HIS and proposes a sparse tensor-based dimensionality reduction (STDR) algorithm which is based on the tensor discriminant analysis and sparse representation. This method firstly applies a group of Gabor filter banks with different scales and orientations on several top components to extract Gabor features which are further rearranged as a second-order feature tensor. A strategy that integrates tensor discriminant locality alignment and sparse representation is used to achieve feature extraction and sparse discriminative information preservation for the second-order tensor. Experimental results obtained with support vector machine (SVM) classifier indicate that this algorithm can obtain the sparse structure description under a tensor framework and maintain the structural integrity of different features with sparse dimensionality reduction which can improve the classification accuracy effectively.Secondly, to deal with the multi-domain united feature extraction and classification problems in HSI, this paper proposes two spectral-spatial classification algorithms:a multi-center SAM-MRF (MSAM-MRF) algorithm and a class-specific random forest (CSRF) algorithm with spectral-spatial hybrid cross-correlation constraints.(1) The MSAM-MRF is based on the spectral angle mapper (SAM) and firstly proposes a multi-center fitting model to overcome the low fitting shortcoming to use the single mean center on behalf of the whole sample set; then, considering that the SAM is a non-probability method, MSAM-MRF makes a Gaussian assumption to introduce MRF into a probabilistic decision framework, obtaining the optimal classification by iterative optimization. The multi-center fitting model urges to partition decision regions more reasonably by decision region splitting with controllable parameters, achieving the deep fitting for the sample set.(2) CSRF is based on the random forest (RF) model and proposes a spectral-spatial hybrid cross-correlation feature extraction method and a classification strategy to build the class-specific decision tree in random forest. The hybrid cross-correlation feature extraction method assumes that all the spectral vectors are the different random variables in a discrete stochastic process, normalizes the relationship between different spectral vectors using the cross-correlation in stochastic process and achieve the construction of spectral-spatial hybrid cross-correlation feature. The classification strategy is obtained by applying the class-specific trees in a random forest model, which the class-specific tree only has one node and makes a binary classification with multi-features fusion decision on the one node.