基于随机森林的WorldVew2i影像建筑物精细提取

Precise Extraction of Buildings' Information in WorldView2 Images Based on Random Forests

针对研究区建筑物大小不一、排列复杂多样、颜色和材质差异较大的实际情况,提出了一种基于面向对象的城区高分辨率影像建筑物信息精细提取方法。该方法考虑了不同颜色建筑物之间以及建筑物与其他地物的特征差异,将建筑物细分为4种子类型,在对高分辨率影像进行分割的基础上,充分挖掘目标对象的光谱、几何、纹理信息等特征,利用随机森林算法对建筑物进行提取并对特征的重要性进行评估。结果发现,精细提取场景下的波段3比值、PCA3均值、PCA4均值、NDVI等特征的重要性较建筑物作为一个类别提取的常规方法出现了较为显著的上升,表明精细提取场景下的影像特征得到了更为充分的应用。使用该方法提取建筑物面积的用户精度和生产者精度较常规方法提高了12.16%和4.09%,为复杂情况下的高分辨率影像建筑物信息提取提供了新的途径。

In view of the diversity of materials and spectra in study area, this article presented a precise extraction method of buildings' information using object-based image analysis. Considering the internal feature differences of buildings, buildings were subdivided into 4 types. And random forests method was used to extract buildings' information and evaluate features' importance. The results show that, the importance of Ratio Layer3, PCA3, PCA4 and NDVI significantly increased in the precise extraction scenarios. It means the feature has been fully utilized. Meanwhile, the producer accuracy and user accuracy of buildings' information are 95.4% and 89.0%, which increase by 12.16% and 4.09% compared with the conventional method. This article provided a new method for buildings' information extraction in complex scenarios.