A hybrid classification method based on fusion of parametric and non-parametric classification algorithms for Landuse/Landcover map in Hirkani Forests

Preparation of proper land use maps has always been one of the important goals of researchers and policymakers. The aim of this study was to provide a new method for preparing land use maps using remotely sensed data and satellite data imagery. For this Purpose, we used Landsat 8 data, Digital Elevation Model (DEM), Principal Component Analysis (PCA), and Spectral Indices to extract land use map in the study area. After all required preprocessing, the training samples were provided. In this study, the training samples were utilized in two parts; in the first part they were used as inputs for image classification using supervised algorithms of maximum likelihood Classification (MLC) and support vector machine (SVM). In the second part, in order to applying Decision Tree Classification (DTC), these training samples were used to determine the spectral reflection of each end-member in the spectrum of electromagnetic waves (image bands, PCA, spectral indices, and DEM).Then, using these binary data and DTC, each end-member was identified and the Landuse/Landcover (LULC) map was extracted. In order to combine the classification results and achieve higher accuracy, the Majority Vote Classification (MVC) method was applied to prepare a new compilation of land use in the area. In order to evaluate the accuracy of produced maps, the statistical parameters extracted from the confusion matrix including overall accuracy, kappa coefficient, user and producer’s accuracy were utilized. According to the results, the combined method (MVC) with a total accuracy of 93.37% and kappa coefficient of 0.91 had the highest accuracy. The overall accuracy of the DTC, SVM, and MLC were 89.61, 88.01 and 87.6%, respectively. Due to the fact that in the nature most of the landuse are mixed and complicated, it would be better to use new methods that cover all aspects of the phenomena. In this research, the data extracted from the supervised classifications as well as the data derived from the DTC were combined and the results clearly illustrate the improvement of the final accuracy of the classification.