以物件為基礎之光達點雲分類

Abstract

近年來隨高解析度衛星影像的發展，影像分類模式由逐像元分類方式逐漸發展為以影像區塊或物件為基礎的分類方法。在以物件為基礎的影像分析方法中，除了額外考量影像物件之光譜、形狀、紋理等特徵作為分類依據之外，亦同步建立像元間的從屬關係，因而提升影像分類的正確性及完整性。為了提高利用光達資料自動分類目標物的能力，本研究將二維物件式影像分類架構應用至三維光達點雲資料的分類。研究方法主要分為三個部分，首先將點雲資料分割為獨立的三維點雲物件。接續利用本研究設計之數種物件特徵，如 Model Ratio、Mean PC、Znormalized、Mean Intensity 等，對各物件進行特徵萃取，以描述點雲物件之空間分佈特性。最後建構分類規則，以決策樹進行點雲物件自動化分類。為驗證本研究於各式光達之適用性，實驗中分別以空載光達與地面光達資料進行自動化地物分類，同時以現有點雲處理軟體LASTOOL與人為方式分類實驗區地物，作為平行化比較及分類品質評估依據。在空載光達部份，研究中選用結構物、樹及車輛作為分類標的，於整體分類精度與Kappa值分別達到98.40 % 與0.9638 之分類成效；在地面光達部份，本研究選用建物、小型結構物、樹、樹幹與樹叢等類別作為分類目標，整體分類精度與Kappa值分別為84.28 % 與0.7221。由實驗成果可知，以物件為基礎之光達點雲分類，能藉由描述點群具有的空間特性輔助點雲判釋，不僅能提升分類成果與人為認知的一致性，在分類品質上亦能有良好的表現。

Recently, with the development of high-resolution images, the image analysis and classification methods have transferred from pixel-based to object-based. Under the consideration of the specific spatial features of objects, such as spectral, shape or texture, or the subordinative relations among objects, object-based image analysis (OBIA) could give assistance to the description of object attributes, which effectively improves the image classification efficiency. In order to raise the capability of automatic recognition of land features from LiDAR data, the 2D object-based classification method is extended for 3D point cloud classification of LiDAR data in this study. The methodology is mainly divided into three parts. First of all, point cloud is segmented to independent 3D objects by various methods. Secondly, in order to describe the spatial characters of these objects, 3D features designed by this study are calculated (e.g., Model Ratio, Znormalized, PFH Index, Difference Ratio). At last, a set of decision rules is built and the point clouds are classified automatically by using the decision tree. To verify the applicability of various LiDAR data, airborne LiDAR and ground-based LiDAR were applied to automatic land feature classification in this study. Meanwhile, LASTOOL, software for LiDAR processing, and manual way were applied to classify the land features in experimental area as bases of parallel comparison and quality assessment. On the part of airborne LiDAR, structures, trees and cars were chosen to be the targets of classification. The overall accuracy and kappa value ran up to 98.40 % and 0.9638 respectively. On the part of ground-based LiDAR, buildings, small structures, trees, trunks and groves were chosen to be the targets. The overall accuracy and kappa value were 84.28 % and 0.7221 respectively. The results show that utilizing the object-based concept to classify LiDAR point cloud can assist point cloud recognition by means of describing the spatial characters of those objects. It then, therefore, improves not only the cognitive consistency of human perception but also the classification quality.