Hyperspectral data in Remote Sensing which have been gathered with efficient spectral resolution (about 10 nanometer) contain a plethora of spectral bands (roughly 200 bands). Since precious information about the spectral features of target materials can be extracted from these data, they have been used exclusively in hyperspectral target detection. One of the problem associated with the detecting process using hyperspectral data is the spectral variation due to topography variability and spectral mixing. Moreover, imperfect sensor noises and atmospheric influences on the target radiance together lead the observed spectral feature of the same material to change in different situations. Target detection methods model the spectral variation in order to compensate their effects on the process. Statistics and subspace based approaches are the two most important methods used in detection process. Statistics and subspace based approaches are the two most important methods used in detection process. Using special statistical assumptions and modeling the spectral variation with limited number of parameters are the main disadvantages of these methods.
One of the strongest detection method is the sparse representation method. It models the differences in the spectral features of targets and background using dictionary matrices. Indeed, it constructs a complete subspace of materials spectrum and their variations.
Building a pure dictionary (clean of spectral mixing) is the main challenge associated in the sparse representation method in the detecting process. Three methods- the dual windows, the global and the learned dictionary- have been introduced in literature. In the dual windows, since it uses outer window to select the target pixels, spectral mixing has not been cleaned.
In the learned dictionary as it uses random picked pixels in order to learn the dictionary, the risk of spectral mixing exists. Furthermore, spectral mixing exists in general method. Considering the disadvantages of the aforementioned methods, in this thesis we introduce a new method to construct the dictionary. Not only do the dictionary atoms provided by this method construct a complete subspace and model spectral variation, but they also are as pure as possible. 
In the proposed method, it is tried to achieve two main purposes which are forming the background subspaces and minimizing the spectral mixing of atoms in the dictionary and target. To this end, correlations between target spectrum and all image pixels are calculated. Afterwards, using image pixels which have different degrees of correlation with target spectrum, different dictionaries are created for the background. Finally, a dictionary is selected from the created dictionaries which presents a complete subspace of image and the subspace also has the lowest correlation with the target spectrum. In this paper, the proposed method of making a dictionary along with a sparsity model, called SRBBH is used and introduced as method Proposed+SRBBH. To survey the efficiency of the proposed methods, a simulation data set and three real data were used, and in order to evaluate the methods, the area below the ROC chart level was used. In experiments performed on both Cuprite and Sandiego data, the area under the graph was 0.9997 and 0.9961, respectively, which shows higher values than other methods. For the other two sets of data, the proposed method performs better than other methods of target detection.