Determination of Best Supervised Classification Algorithm for Land Use Maps using Satellite Images (Case Study: Baft, Kerman Province, Iran)
Authors
Abstract
According to the fundamental goal of remote sensing technology, the image classification of desired sensors can be introduced as the most important part of satellite image interpretation. There exist various algorithms in relation to the supervised land use classification that the most pertinent one should be determined. Therefore, this study has been conducted to determine the best and most suitable method of supervised classification for preparing the land use maps involving no grazing, heavy and moderate grazing rangelands, ploughed rangelands for harvesting licorice roots and dry land and fallow lands in Baft, Kerman province, Iran. After being assured of accuracy and lack of geometric and radiometric errors, the images of Landsat and ETM+ sensors achieved on 3 July 2014 have been used. A variety of algorithms involving Mahalanobis distance, Minimum distance, Parallelepiped, Neural network, Binary encoding and Maximum likelihood was investigated based on field data which were obtained simultaneously. These algorithms were compared with respect to error matrix indices, Kappa coefficient, total accuracy, user accuracy and producer accuracy of maps using ENVI 4,5. The results indicated that the Maximum likelihood algorithm with Kappa coefficient and total accuracy of map estimated as 0.969 and 97.77% were regarded as the best supervised classification algorithm in order to prepare the land use maps. Mahalanobis distance algorithm had a low ability for recognizing two types of dry land and fallow land uses concerning the extracted maps. According to the findings, various land use maps as rangelands under three grazing intensities and ploughed rangelands to harvest the licorice roots provided by the means of algorithms related to neural networks were not of sufficient accuracy. The highest Kappa coefficient of Neural network algorithms was estimated as 0.5 and attributed to the algorithm of multilayer perceptron neural network with the logistic activation function and one hidden layer.
