Comparison of three common methods in supervised classification of satellite data for vegetation studies

Usage of modern technologies such as GIS and RS in plant ecosystems studies and especially land cover mapping is needed to recognition of these instruments efficiency and also identification of the best methods for applying them. This study aimed to compare the efficiency of three common supervised classification methods of satellite data (Minimum to Distance, Parallelepiped and Maximum Likelihood) to identification of plant groups in Goloul-via-Sarani protected area, Northern Khorasan Province, Iran. In order to this, 143 training samples (>30m2) were collected from areas that shown a homogenous plant species composition. These data recorded by GPS device and so transported to a GIS database. Satellite data included Landsat ETM+, and IRS-P6 LISS III that were prepared and analyzed by ENVI 4.2 software. Amount of efficiency for each method was evaluated by measurement of overall accuracy (OA) and Kappa coefficient (KC) criteria. Results showed that ML method makes the highest accuracy for two data series (OA=90.35, 82.19 and KC=0.878, 0.772 for Landsat and IRS data respectively). In the face, PP method showed the worst results (OA=67.09, 58.76 and KC=0.593, 0.478). These results suggested that collection of sufficient training samples from natural classes and surveying probability of image pixel's dependency on these classes can be useful for classification of plant groups.