Abstract: The classification of remote sensing image is an important part of extracting effective information of images. In order to explore the optimal classification methods, many machine learning algorithms are gradually applied to the classification of remote sensing images. Because of the high efficiency, speediness, and good performance generalization, extreme learning machines (ELM) have been widely used in pattern recognition. This paper aims to classify high-resolution remote sensing images, and analyze the performance of extreme learning machine algorithms for the accuracy of classification of remote sensing images, and compare ELM algorithm with support vector machines (SVM) algorithm and Maximum likelihood method. The GF-2 data from some area in Changchun City were selected to test the accuracy of all the three methods for classification with the fused image as the original data. The results show that the overall accuracy of ELM algorithm is more than 85%, and the kappa coefficient is 0.718. Compared with the other two classification methods, the ELM classification accuracy is the best, and its running time is faster than the support vector machine by 2 480 s, 1/8 of the support vector machine's. An even better performance can be obtained in a shorter training time. ELM is valuable for classification of remote sensing images.

Key words: extreme learning machine, remote sensing image classification, GF-2 image, supervised classification, support vector machines