COMPARISON OF ACCURACY ASSESSMENT FOR SOFT CLASSIFICATION TECHNIQUES

Abstract

Remotely Sensed data is an ideal source of data for mapping land cover and land uses at variety of spatial and temporal scales. Remotely sensed images often contain a combination of both pure and mixed pixels. The easiest and usual hard classification techniques assign mixed pixels to the class with the highest proportion of coverage or probability. Loss of information is inevitable during this process. For this reason it has been proposed that fuzziness should be accommodated in classification procedure so that pixels may have multiple or partial class membership. To preserve the information present inside a pixel Soft classification techniques were introduced: they assign pixelfractions to the land cover classes corresponding to the represented area inside .a pixel. In this research work statistical learning based classifier Support Vector Machines (SVM) with different kernels and combination of kernels has been used. In this research work density estimation based on the support vector machines (SVM) approach, uses Mean Field (MF) theory for developing and an easy and efficient learning procedure for the SVM. The MF methods provide efficient approximations, which are cope with the complexity of probabilistic data models. In this research work, a full fuzzy concept has been tried, at sub-pixel level, using density • estimation based on support vector machines ,(SVM) approach. This approach was evaluated with respect to fuzzy weighted matrix. The tool required. to achieve the objectives mentioned in this work were not present in • commercial digital image processing software's. For this research work, software package in JAVA environment known as SMIC: Sub-Pixel Multispectral Image classifier had been used. This software package has file, signature, classifier and accuracy assessment model required in this research work. In this study, sub-pixel classification approach has been carried out using data set from LISS-III • and LISS-IV sensors of IRS-P6 , (Resourcesat-1). Fraction Image generated from different classifiers had been evaluated using FERM and SCM. From different experiments carried out in this thesis we conclude the following results: Among different single kernels in experiment (a) KMOD kernel of SVM give us highest (82.25 (%) using FERM- and 83.71±14.23 (%) using SCM) overall accuracy. From experiment (b) we conclude that KMOD kernel gives better result compare to RBF -kernel. The Average User's accuracy for RBF and KMOD kernels are 72.71 and 87.71 using -FERM respectively and the average Producer's accuracy for RBF and KMOD kernels are 66.08 and 78.22 using FERM respectively. Experiment (c) shows that overall accuracy of soft-classification has improved significantly when it is mixed with spectral this is due to the robustness of spectral kernel for illumination. From this experiment we can observe that overall soft-classification accuracy for KMOD-Spectral kernel is highest (90.59 from FERM and 91.59±7.06 from SCM). Sigmoid Spectral kernel has Second highest overall accuracy (78.81 from FERM and 80.76±9.14 using SCM). Experiment (d) shows that mixing of Gaussian kernel with Inverse Multiquadric gives highest overall accuracy (84.22 using FERM and 86:01±3.52 using SCM