Abstract: In order to quickly and accurately grasp the spatial distribution and dynamic change information of impervious surface, based on the idea of multi-classifier ensemble learning, random forest algorithm is introduced in eliis paper. Landsat 8 image is used as data source and Changchun city as experimental area. 25 feature variables, such as spectral indices, texture measures and independent components after spatial transformation are selected to classify. The importance of variables calculated by out of bag error is analyzed and the optimal classification model is obtained through many experiments. The extraction of high-precision impervious surface is also realized. Finally, random forest algorithm is compared with the traditional parameter classifier. The result indicated that the overall accuracy of random forest algorithm can reach 94%, which is higher than 5.9% of maximum likelihood classification, 0.77% of support vector machine algorithm, 0.914 3 of Kappa coefficient and 0.104 3 of root mean square error. The extraction accuracy of impervious surface is 95.54%, which can not only accurately extract impervious surface but also provide effective thematic information for urban construction and planning.

Key words: impervious surface, random forest, image classification, feature extraction