Clustering accuracy analysis of building area in high spatial resolution remote sensing images based on k-means algorithm

Abstract

Buildings and adjacent objects in the high spatial resolution images Present the spatial correlation due to the spectral similarity. In addition, the spectral details of building top are completely reflected in images because of resolution levels increased from meter to sub-meter. K-means algorithm is a classical clustering algorithm. Fine spectrum, low signal to noise ratio (SNR) and high spatial heterogeneity of high spatial resolution image (K-Means) pose a great challenge to the accuracy and precision of clustering segmentation. The contents and sizes of the clustering scale window determine the setting of the parameters like classification numbers, the size of calculating data sets and the change of clusters in clustering before running the algorithm. It is very important to improve the clustering accuracy by deeply studying the change of clustering accuracy caused by the change of window contents and size. In this study, Geoeye and QuickBird high spatial resolution images were selected as experimental images. In which a large number of sub-windows of different building pixel proportions are selected(1%∼70%). The factors of separability and building density were used for describing image classification characteristics of clustering window. The sensitivity factor of K-Means clustering accuracy is studied from the viewpoint of spectral separability of high spatial resolution images target and background. The linear and nonlinear function relationship were established between the precision index of global and local window overlapped area and the characteristic factors of image target. Finally, the classification guiding experience that is not exactly same was obtained under the formula derivation. It provides an effective reference and basis for high spatial resolution remote sensing image segmentation and classification quality assessment.