高分辨率影像中城区树冠多尺度聚类识别方法

Multi-scale method of urban tree canopy clustering recognition in high-resolution images

提出一种约束均值漂移方法,对高分辨率影像中的城区树冠进行提取。该方法首先进行小波分解,建立小波金字塔结构,用特定窗口,对每一层小波的低频系数计算均值,同时对其高频系数计算标准偏差,在每一层,用这些均值和标准偏差构成特征空间,最终构成多尺度金字塔影像特征空间;然后,从金字塔顶层开始,逐层进行均值漂移计算,并在层间进行尺度传递,由于尺度传递可能造成特征空间更加不平滑,所以本文采用约束均值漂移方法进行聚类,实现城区树冠初步聚类分割。最后,由于特征空间的特征可区分性很难保证在区域边缘处的聚类精确性,所以本文进一步采用基于聚类特征的监督分割方法提取树冠。实验结果表明,与传统的直接监督方法以及非监督方法相比,该方法能较好地消除高分辨率导致的影像高度细节化等因素对城区树冠提取的影响,具有很强的实用性。

A constrained mean shift method for extracting urban tree canopy of high-resolution images is presented. First, a wavelet is decomposed and a layered pyramid structure is established. Using a specific window, the mean of the low-frequency coefficient and the standard deviation of the highfrequency coefficient of each wavelet layer are computed. The computed mean and standard deviation are used to constituted. top layer of constitute a feature space in each layer a multi-scale pyramid image feature space is Second, from the top of the pyramid, the mean shift of each layer is computed from the the pyramid, and the scale transfer between layers is carried out. The scale transfer may cause the feature space even more unsmooth, so a constrained mean shift method is adopted to realize preliminary urban tree canopy clustering segmentation. Finally, as the distinction of features in a feature space is difficult to guarantee the clustering accuracy at the edge, a further supervised segmentation method based on clustering features is taken to extract the tree canopy. Experiment results demonstrate that compared with traditional supervised methods and unsupervised methods, theproposed method can eliminate the effects of over-detailed images and other factors caused by highresolution on extracting urban tree canopy.