A comparison of pixel-based and object-based approaches for land use land cover classification in semi-arid areas, Sudan

Abstract

This paper deals with the comparison between application of pixel-based and object- based approaches in land use land cover classification in semi-arid areas in Sudan. The second aim is to assess the accuracy of classification for each approach. The study was conducted in the gum arabic belt in North Kordofan State, which is affected by modifications in conditions and composition of vegetation cover trends. The study used ASTER L1B registered radiance at the sensor image acquired on (19.10.2010). The image was radiometrically corrected by using ENVI-FLAASH software. Subset with an area of (40880) ha was created. The image classification (pixel-based and object-based) and accuracy assessment were conducted. Total number of (47) GCPs were surveyed and used in accuracy assessment using ERDAS 9.1. Image segmentation process was implemented using Definiens eCognition 7.1 software. Segmentation level 4 of scale parameter 25 was selected for classification based on colour and form homogeneity. Land use land cover classes were derived by classification using the nearest neighbor classifier with membership functions (fuzzy logic) for each class. The land use land cover distribution in the area for forest dominated by Acacia Senegal is (20%) and for residential area is (1.50%) for the two methods of classification. While for bare and farm land, grass and bush land and mixed woodland classes are (6.69% and 1.63%), (18.62% and 15.16%) and (53% and 61%) for pixel based and object based methods, respectively. The overall accuracy and Kappa statistic of the classification produced by the pixel-based and object-based were (72.92%, and 54.17%) and (0.6259 and 0.3810), respectively. The pixel based approach performed slightly better than the object-based approach in land use land cover classification in the semi-arid land in gum Arabic belt.