STUDY AND IMPLEMENTATION OF SOME.PER-PIXEL CLASSIFIERS FOR HYPERSPECTRAL IMAGING

Abstract

Remotely sensed images are attractive sources for extracting land use and land cover information, where an image classification algorithm is employed to retrieve land use and land cover information. The images may be panchromatic, multispectral or hyperspectraI depending upon the number of bands. Existing statistical classifiers have shown marked limitations for hyperspectral data. Statistical classifiers are very much 'dependent on statistical parameters. Therefore, these classifiers are not suitable for hyperspectral data. To overcome these limitations, some new age non-parametric algorithms came into existence. These include support vector machines, artificial neural network algorithms, evidential reasoning, decision tree classifier and many others. However, each has its own limitations. Nevertheless, an effective algorithm for image classification should be computationally efficient. In this dissertation, two algorithms namely Support Vector Machines (SVM) and Spectral Correlation Mapper (SCM) have been implemented for per pixel classification from hyperspectral data. GUI based software has been developed to implement the two classifiers. The performance of these classifiers has been assessed through a set of experiments conducted on synthetic as well as real remote, sensing dataset obtained from hyperspectral sensors. The accuracy of these algorithms has been assessed using our own developed accuracy assessment module. The impact of factors such as the choice of the multiclass method, the optimizer and kernel function on SVM classification has also been studied. The efficacy of feature extraction as a pre-processing step to SVM classification to reduce the dimensionality of the data has also been examined. The - results showed that feature extraction did not yield any significant advantage whereas selection of training sample size was an important consideration for the successful implementation of any classifier. Therefore, the sensitivity of SVM based classifier with respect to training sample size was also investigated. . The SVM classifier performed quite well with the use of small training sample size. The linear and polynomial degree two kernels exhibit superior performance than the other kernels. Finding out exact range of penalty values for the kernel functions, using trial and error is a difficult task in case of SVM classifier. On the contrary the working principle of SCM is quite different from SVM algorithms. It depends upon the spectral correlation of the data sets. First the SCM algorithm applied for image ii classification purpose. For AVIRIS dataset 34% accuracy has been achieved whereas for HYPERION dataset corresponding value is 82%. SCM has also been applied for target detection purpose in some synthetic and real data sets. The targets have been identified properly for most of the cases. Further, a ROC (receiver operation characteristics) analysis has been carried out for a performance assessment of spectral correlation mapper. It shows more than 98% accuracy for ROC analysis.