Multi-scale urban land cover extraction based on object oriented analysis

The paper classifies urban land cover object-oriented. First, use the Mean Shift algorithm to segment the digital aerial image of study area. By changing the algorithm initial-parameters, different scales of the segmentation results are got, and an optimal scale segmentation result is selected from these results, as the data source of the classification process. Then, by analyzing the spectral features, texture features, as well as the topographical features of the study area, the Fisher criterion is taken to calculate the classification ability of each feature and sort them in descending order. In order to avoid the “Hughes phenomenon”, tests are taken to find the optimal feature space dimension. Finally, by using ensemble learning algorithm, decision tree algorithm, which is a weak learner will be upgraded to strong learner to improve the classification accuracy, and then the rule which is produced by the strong learner is used to classify the study area. The classification accuracy is 89.87%, which means that the method can effectively carry out urban land use/cover information extraction; furthermore, because these algorithms used in this article are no limitation in scale, so they are also suitable for multi-scale remote sensing image classification.