Comparison of Pixel and Object Based Post-Classification Algorithms for Monitoring Wetland: A Case Study in River Delta.
Vakaki, Eleni1; Tsakiri-Strati, Maria2; Doxani, Georgia2; Siachalou, Sofia2; Mallinis, George3
1Aristotle University of Thessaloniki, GREECE; 2AUTH, GREECE; 3DUTH, GREECE

Ecosystems are in a permanent flux at a variety of spatial and temporal scales all around the world (Coppin et al. 2004). Anthropogenic activities influence the natural environment to such an extent, so it is imperative that changes are being monitored at a high temporal scale. Especially as far as fragile ecosystems are concerned, change detection algorithms are needed as a part of efficient environmental strategies. For this reason various researchers have focused on improving the accuracy of change detection methods (Singh 1989, Ridd and Liu 1998, Mas 1999, Liu et al. 2004). The aim of this study is to compare and evaluate object-based with pixel-based post-classification techniques. The study area is the endangered wetland ecosystem of Nestos river Delta, in Northern Greece which is protected by the Ramsar convention and Natura 2000.

In this study, two Landsat images were used dating from 1995 and 2007. The first step of the study was the geometric and radiometric processing of the images. It is worth mentioning that the Landsat TM image of 1995 had partial cloud coverage, so it was necessary to apply an algorithm to automatically detect and remove cloud and its shadow (Song and Civco 2002, Irlsh et al. 2006, Martinuzzi et al. 2007). The cloud/shadow classification was performed using an object based algorithm after selecting certain object-features (mean values of bands 1, 2, 3 and 4).

After the pre-processing, the images were classified with pixel based and object based classification. In order to collect the training data, two VHR images from the same dates, were used as ground truth. A pixel based supervised classification was implemented using the nearest neighbor method. The selected classes were: sea, inland waters, salt marshes, cultivations, broad leaved forest, beaches-dunes-sands, transitional woodland and shrub.

For the object based classification algorithm, the images were segmented and then classified using different features of each image such as the values of the bands, the Normalized Difference Vegetation Index, the texture, the brightness of the image and class related features. After classifying the images with both methods, the post-classification algorithm was applied to detect the land cover changes between the two dates. The final purpose of this study was to compare the results of both methodologies. This step involved the estimation of the error matrix and the Kappa index for each change image. The result of the accuracy assessment highlighted the prevalence of the object-based post-classification technique compared with the pixel-based technique.