On the use of binary partition trees for the tree crown segmentation of tropical rainforest hyperspectral images
On the use of binary partition trees for the tree crown segmentation of tropical rainforest hyperspectral images
复制标题
DOI：
10.1016/j.rse.2014.12.020
复制
发表时间：
2015-03
影响因子：
13.5
通讯作者：
G. Tochon;J. Féret;S. Valero;R. Martin;D. Knapp;P. Salembier;J. Chanussot;G. Asner
中科院分区：
文献类型：
--
作者：
G. Tochon;J. Féret;S. Valero;R. Martin;D. Knapp;P. Salembier;J. Chanussot;G. Asner
热带森林上空获取的遥感图像的分割对于许多生态应用非常重要，例如森林清查或生态系统的养护和管理，对于这些应用，物种分类技术和个体数量的估计是非常有价值的投入。本文提出了一种基于二叉分割树(BPT)算法的高光谱图像分割方法，并将其应用于夏威夷热带雨林和巴拿马热带雨林的两个地点。在构建BPT之前，比较了结合空间和频谱降维的不同策略。采用分水岭变换、均值漂移聚类等多种超像素生成方法对图像进行空间降维，得到初始分割图。采用主成分分析方法对光谱进行降维，并对不同的主成分组合进行比较。提出了一种基于直方图的非参数区域模型，并结合扩散距离进行区域合并。提出了一种基于合并区域大小不连续的自适应剪枝策略，并与已有的剪枝策略进行了比较。最后，介绍了一套评价树分割质量的标准。该方法正确地分割了高达68%的树冠，并生成了合理的分割景观格局。
The segmentation of remotely sensed images acquired over tropical forests is of great interest for numerous ecological applications, such as forest inventories or conservation and management of ecosystems, for which species classification techniques and estimation of the number of individuals are highly valuable inputs. In this paper, we propose a method for hyperspectral image segmentation, based on the binary partition tree (BPT) algorithm, and we apply it to two sites located in Hawaiian and Panamean tropical rainforests. Different strategies combining spatial and spectral dimensionality reduction are compared prior to the construction of the BPT. Various superpixel generation methods including watershed transformation and mean shift clustering are applied to decrease spatial dimensionality and provide an initial segmentation map. Principal component analysis is performed to reduce the spectral dimensionality and different combinations of principal components are compared. A non-parametric region model based on histograms, combined with the diffusion distance to merge regions, is used to build the BPT. An adapted pruning strategy based on the size discontinuity of the merging regions is proposed and compared with an already existing pruning strategy. Finally, a set of criteria to assess the quality of the tree segmentation is introduced. The proposed method correctly segmented up to 68% of the tree crowns and produced reasonable patterns of the segmented landscapes.