Experimental study on graph-based image segmentation methods in the classification of satellite images.

Object recognition is one of the primary tasks in remote sensing. For example, identifying land cover based on satellite images has an important role in agriculture, environmental protection and economics. Image segmentation is an optional elementary step of the classification process. It can improve both accuracy and performance. Graph theory is a powerful tool to describe image processing algorithms. Its theoretical results greatly help in the analysis of methods. In this article four graph-based image segmentation algorithms are compared and evaluated, namely the best merge algorithm of Beaulieu, Goldberg and Tilton, tree merge segmentation of Felzenszwalb, minimum mean cut segmentation of Wang and Siskind, and finally normalised cut algorithm of Shi and Malik. After segmentation, segments are assigned to land cover categories with supervised classification. In turn, the result of classification is used to measure the accuracy of the procedure. Authors will describe the theoretical background and implementation details of segmentation algorithms, and will introduce some possible improvements.