A Novel Sparse Based Classification Approach For Hyperspectral Data

The classifcation of high dimensional data is a difcult problem in pattern classif-cation. In remote sensing, the problem is particularly serious because of difculty, costand efort involved in assigning labels to the data. In hyperspectral data classifcation,the challenge is even more pronounced. Hughes phenomenon[1]states that to get thehigher classifcation accuracy of high dimensional data, the number of labeled samplesmust be increased. In comparison to earlier remote sensing technologies, hyperspectraldata is more high dimensional and, therefore, more labeled samples are required to learnthe model in order to get good classifcation accuracy. This research work focuses ona major challenge of classifcation of hyperspectral data using a few labeled samples inoriginal dimensional space, as it would reduce the need to have large number of labeledsamples, and eventually, would help in reducing the overall cost and efort required forthe classifcation applications.The most commonly used approach is to use dimension reduction to reduce the di-mensions of hyperspectral data in order to deal with few training samples. This researchwork proposes a dimension reduction technique to deal with the major challenge of fewlabeled samples. But, the use of dimension reduction is debatable, and contradicts theoriginal purpose of having hyperspectral imaging technology, because it reduces the in-formation, after all. On the contrary, the primary purpose of hyperspectral imaging isto provide reflective information in more wavelength bands and, therefore, to aid in thediscrimination of many materials on the surface of earth than which can be provided bythe earlier remote sensing technologies. This research work, further, focuses on a clas-sical pattern classifcation approach using adaboosting of artifcial neural network basedweak classifers, and the results are promising. But, there are certain drawbacks as well,i.e., a lot of labeled samples are needed for learning purposes, and learning is time con-suming as well. To deal with all of the above mentioned challenges in hyperspectral dataclassifcation, this research work fnally proposes a sparse based classifcation approach,and claims the sparsity of hyperspectral data. This subsequently, is a fnding of this re-search work after the exploration of the properties of hyperspectral data. It is a furtherfnding that a few labeled samples can be used to classify the data in original dimensional space. The sparsity of hyperspectral data is exploited using1-minimization based sparserepresentation for the classifcation purposes. This work assumes that the data withineach hyperspectral data class lies in a very low dimensional subspace. Unlike tradition-al supervised methods, the proposed method does not have separate training and testingphases and, therefore, does not need a training procedure for model creation. Further,to prove the sparsity of hyperspectral data, and handle the computational intensivenessand time demand of general-purpose LP-solvers, this work proposes a Homotopy basedsparse classifcation approach, which works efciently when data is highly sparse. Theapproach is not only time efcient, but it also produces results, which are comparable tothe traditional methods. Extensive experiments prove that hyperspectral data is highlysparse in nature, and the proposed approaches are robust across diferent databases, ofermore classifcation accuracy, and are more efcient than state-of-the-art methods.